LMM — Living Memory Model

A verifiable memory layer for language models. Facts live in a graph, every claim carries its source, and an answer the memory does not support cannot leave the system.

📚 Documentation — concepts, guides, the full API surface, and every measurement with its receipts.

📄 Scaling Memory Instead of Scaling Parameters — why this exists, measured against GraphRAG, and the two bugs I found in my own scorer.

In an LLM, knowledge is frozen into weights at training time. In LMM, knowledge lives in a memory you can write to, inspect and audit — which is what makes "I don't know" trustworthy rather than polite.

pip install living-memory-model      # the import is `lmm`

from lmm import Memory

m = Memory("mind.lmm")            # persistent; Memory() is transient
m.learn("manual.pdf")             # pdf · docx · xlsx · csv · md · txt · or text
print(m.ask("what is the screen's diagonal?"))
m.save()

That is the whole API. One method reads every format, and the file extension picks the adapter.

---

Try it without changing anything

The smallest useful piece of this library is a function over two strings. It needs no Memory, no ingestion, no engine call and no network — keep your retriever, your vector store and your model:

from lmm import check

v = check(answer, "\n".join(chunks), question)
if not v.ok:
    print("not carried by the evidence:", v.unsupported)

The valve opens at 8 bar.                 ok=True
The valve opens at 9 bar.                 ok=False   unsupported=['9']
The valve opens at 8 bar in the cellar.   ok=False   unsupported=['cellar']

It returns ok, figures_ok, coverage, unsupported and kept — the sentences of the answer the evidence does carry.

It is a pre-filter, not a faithfulness judgement, and the difference matters. These are deterministic lexical readings: a paraphrase the evidence does carry comes back not-carried, and a sentence spliced from two lines the document never put together comes back carried. It says nothing about whether the answer is RELEVANT to the question — a true, sourced sentence about something else passes. Inside the library those cases are what the engine-side read-back and relation check are for; this is the free tier, and those are the price of free.

---

How it works

The gate is drawn twice on purpose. On the way in it decides what may become a record; on the way out it re-reads the sentence the engine produced and drops any claim the memory does not support. Everything else in this README — the abstention rate, the audit trail, the cost — follows from those two edges.

Two details worth naming. A spreadsheet reaches the graph with no model call at all, because a table already states its own structure; that is why a sheet loads in milliseconds. And the engine is drawn smaller than the memory because it is smaller: it phrases, it does not know.

Three channels, all local, all explainable

Retrieval reads the same store three ways, and each can say why a line arrived:

channel	what it reads	the statistic
lexical	the words a line carries	information content: log(1+N/df), and log(S/s) for a word's power to separate documents
structural	records, fields, and the entities a document repeats	a field is a head the corpus repeats; an entity is a phrase whose words occur together beyond chance (PMI), keep varied company, and do not choose freely
meaning	which DOCUMENT a question is about, and then the order of the lines inside it	a bundled static multilingual embedding: one lookup per word-piece and a mean, fused with the words by reciprocal rank (RRF, K=60), then reordered by late interaction

None of them calls a model, none of them needs a vendor, and none of them widens what may be SAID: they widen what can be FOUND, and the gate reads the evidence exactly as before.

The meaning channel ships switched on. A reader who paraphrases — a lease that says RESIDES against a question that says LIVES — is the case the words cannot reach, and an optional extra is a feature most installations never turn on. So the vectors are IN the package: pip install living-memory-model, and it works offline with nothing to configure. It is a matrix, not a model (one lookup per piece and a mean), so there is no torch and no warm-up. The one dependency this buys is numpy, and it is named in pyproject.toml rather than hidden.

Measured on this repository's field corpus — 20,000 lines, 180 known-item queries with each query's two most distinctive terms removed:

encoder	first hit	MRR	index, 20k lines
a distillation of our own	27%	0.306	1.2 s
paraphrase-multilingual-MiniLM-L12-v2	66%	0.710	36.1 s
the bundled matrix	84%	0.883	0.3 s

And end to end, by whether the ANSWERING line reaches the five the engine is shown (115,913 lines, 304 queries):

	first	in five
words alone	31%	32%
words + meaning channel	31%	60%
...and reordered by late interaction	34%	79%

The ceiling stays open — Memory(encoder=...) takes any callable from strings to vectors, reranker= any callable from (query, texts) to an order — and the floor is safe: dense=False, a missing numpy, or an encoder that raises leaves the channel absent, never degraded, and the words carry the turn.

Which layer the meaning channel sits at was measured, not assumed. Over lines it moved nothing; over DOCUMENTS it took the store's own retrieval from 92% to 99% on 103 known-item queries. A catalogue line reads "Aidiyet ve Motivasyon." — four words carry no subject — so vectors over lines compare fragments while vectors over a document profile compare subjects. The opposite arrangement was measured too: choosing the lines INSIDE a proposed document by meaning rather than by words took the answering line from 78% to 31%. Meaning says which document; inside it, the question's own words are sharper.

The entity channel is also the partition. Measured on a 36,472-line novel: asked what connects two entities, the lexical search finds a line carrying both — six pairs out of six — and takes 417 ms doing it, because every line carrying either name is scored. The same lines come out of two posting lists in 0.01 ms. Not a better reading; the same reading without reading the document.

---

The composer: a plan is a proposal, the primitives are the law

A memory that answers only the question shapes somebody wired by hand answers a finite set of questions, and shapes are endless. So the shapes are composed rather than coded:

question ──► the engine proposes a PLAN over verified primitives
             a = anchor: the recital        (its dated line, read from the
             b = anchor: the gala            sentence before the envelope)
             out = span: a, b               (arithmetic, ours)
                    │
                    ▼
             the interpreter executes ONLY operations it knows, on phrases
             the STORE can anchor — an unknown operation, an unanchorable
             phrase, a dangling reference: the plan dies, no claim is born
                    │
                    ▼
             the sentence is built from the final step's TYPED value by our
             template; the engine contributes operation names and phrases,
             never an output word

The primitives are the measured organs: anchor a phrase to its dated line, take the latest, gather lines, span two dates, order them, filter by a cutoff, count, tally by month. Measured on a dated store, five question shapes nobody wrote an organ for — did A happen before B, in which month most, how many days between, when first, when last — answered correctly with no new code.

And the seat binds the plan's output TYPE: a count-shaped turn will not voice a month tally, however correct that tally's arithmetic. Organs silent and plan dead, the turn refuses rather than gambles the prose chain — measured on a benchmark slice where the chain earned two right answers and twenty-three wrong ones.

Dialogue is data with two more columns

Chat is trickle data: events melt into talk, and two things every conversation carries were being flattened into prose.

When it happened is the sentence's word before the envelope's. People tell events days later — [stamped April 10] back on March 22 the GPS died — so the engine reads the candidate lines and proposes the date; the arithmetic decides whether the reading may stand (the date must equal some candidate's stamp, or its day must be written as a number in that line with the year in the stamp's neighbourhood). An invented date fails both and the envelope speaks, as before. No month table, no date grammar.

Who spoke is metadata as real as the date. learn(..., speaker=) keeps it per line, session.asker says who is asking, and the event organs seat that speaker's lines first — nothing excluded, priority alone moves.

When the question is asked is metadata too (0.7): session.asked_at takes a datetime.date, the plan's now primitive reads it, and "how many days ago…" becomes arithmetic between a line's stamp and the asking day. With none set, now is the calendar's today — a benchmark replays last year's questions, a letter is answered a week late.

m.distil(text) is the write-time reading for passages where events melt into talk: the engine LISTS each thing as thing/what-happened/KIND — ongoing facts included, "I'm also getting X" is no less a fact than "I bought X" — the passage's own words admit the thing, and each survivor becomes one gated dated record. The KIND ("magazine subscription") is an index key beside the bridges, never a claim: it widens what a later question can FIND and no gate or spoken word ever reads it. The passage is kept as evidence, so nothing is lost. Documents that state their own structure need none of it; their rows reach the graph for free.

---

Why not embedding RAG

	Embedding RAG	LMM
Ingestion	chunks + embeddings	facts into a graph + sentences into an evidence index
Retrieval	similarity gamble	deterministic word/graph lookup, explainable
Multi-hop	fails when chunks don't co-retrieve	derives new facts symbolically (µs, no model call)
Paraphrase	the embedding's strength	the open gap — a bridge learned per field, and the engine's second ask; a dense channel is the instrument this wants and is not built yet
Entity graph	built by an LLM (GraphRAG), a claim	built by the document: an entity is a collocation, an edge is a witnessed co-mention carrying its sentence
Fabrication	a plea in the prompt	structural gate: unsupported claims cannot leave
Provenance	none	every fact carries its source; uncertain answers are flagged
Tables	flattened into text	rows go straight into the graph — zero model calls
A 350-page manual	seconds of embedding, lossy retrieval	queryable in ~2 s

The core install has one dependency, numpy. The graph, the gate, the trust ordering and the derivation are pure python — no model, no network, no GPU — and numpy is there for one thing: the meaning channel, a 30 MB multilingual matrix that ships in the wheel so a paraphrased question is understood out of the box. Memory(dense=False) runs without either.

Measured

Same questions to both sides. RAG baseline: LangChain + multilingual embeddings

• Chroma + gpt-4o-mini, best-practice grounding prompt. LMM's engine here is the same gpt-4o-mini, so the column compares architectures, not model sizes. LMM figures are medians of three samples — these benchmarks are noisy and a single run is not a measurement.

Benchmark	RAG	LMM
Fictional corpus, 17 q (no parametric leakage possible)	14/17	17/17
A strategy document, 16 q	13/16	14/16
A 350-page device manual, 30 q	14/30	27/30 — ingested in 2.3 s vs 9.2 s
An inspection spreadsheet, 15 q	15/15	14/15 — ingested in 86 ms vs 4.7 s
Wrong facts asserted, all runs	1	0

The bigger the document, the wider the gap: retrieval collapse is RAG's structural ceiling, composition is LMM's structural strength.

Four corpora, one code path

The table above is one document at a time. These are the sets the answer path is developed against — three runs each, medians, same engine:

Corpus	factual	comparisons	frontier*	traps (fabrication)
62 sibling training outlines, 73 q (customer's, not redistributable)	40/40	15/15	7/8	10/10
12 hardware specifications, 19 q — a corpus the system had never seen, no adaptation	8/8	4/4	3/3	4/4
1,000 generated specifications, 50 q (questions derived from the corpus, not written by hand)	25/25	—	—	25/25
NIST SP 800-63B (public PDF), 13 q · fictional EN corpus, 17 q	11/13 · 17/17	—	—	—

Every category above is a median of three runs with zero flips — no question changed its verdict between runs — except the single frontier miss.

* frontier: questions whose discriminating word the documents never write ("the priciest" against a column headed LIST PRICE).

Against three memory frameworks, same corpus, same engine

LangChain (RAG), LlamaIndex and Mem0 answering the same questions over the same 62 documents, each with Azure gpt-4o-mini behind it, two repeats:

	factual	comparisons	frontier	traps	wrong facts asserted	p50 latency	ingestion
LangChain	21/40	14/15	4/8	9/10	1	1.9 s	75.6 s
LlamaIndex	29/40	13/15	6/8	9/10	1	3.4 s	12.9 s
Mem0	25/40	15/15	5/8	8/10	2	2.2 s	161 s
LMM	40/40	15/15	7/8	10/10	0	5.9 s	2.1 s

The shape of that table is the design: LMM is ahead on every accuracy column and alone at zero fabrications, loads a corpus in seconds rather than minutes, and pays for its verification in per-question latency — three model calls to answer, three more to audit the answer, where a RAG pipeline makes one and asserts whatever comes back. The questions it can answer from a record cost nothing at all (0.01 s), which is what pulled the median down from 11.2 s.

Speed. A question that names a document and one of the corpus's own field heads is answered from the record itself — no model call, no composition, the document's line and its stamp: 0.01 s where the same question used to take 10.4 s and twelve calls. It fires on 33 of the 40 factual questions above and on none of the comparisons, frontier questions or traps, which is the shape it should have. Questions that need the engine still cost what verification costs: 8 s for a comparison, 7 for an honest refusal.

Scale, measured on the 1,000-document store (25k sentences): ingestion 2.2 s, 82 MB resident, retrieval 121 ms median with the named document's line recalled 50/50. Conversation depth: the same twelve questions scored 11/12 asked to a fresh session and 11/12 asked after twenty-four turns of other documents, pointers and topic switches — the same single failure in both.

Only the invented corpus is in this repository. The other three documents belong to their owners — a manufacturer's product manual, an institution's internal strategy document, an inspection report naming real people — and are not ours to redistribute. benchmarks/corpus*.txt is a fictional world written so that no answer can come from what a model already knows.

These numbers are lower than this file used to show, and that is the point. An audit removed everything that had been fitted to the documents being scored: few-shot examples quoting the manual's own voltages, a prompt clause teaching the exact distinction one question turned on, and thresholds (a 700-character window, a 40-character cell, a 0.75 similarity) chosen while reading one PDF. The honest cost was 62/63 → 59/63, of which the prompt content alone was 4 points — that was the share of the score that came from having seen the test. Removing the last Turkish from the runtime cost one point more.

Against GraphRAG, measured

Beating chunk-and-embed is the obvious comparison. The one that matters is against graph-structured memory, so this runs Microsoft's own package — pip install graphrag, its indexer, its prompts, its defaults, local search — on the same corpora, with the same engine, scored by the same side-blind evaluator. Nothing is reimplemented, so nothing here is a strawman. benchmarks/graphrag_side.py.

Same engine (gpt-4o-mini), 3 samples each	LMM	GraphRAG
Fictional corpus (1.6 KB), 17 q	17/17 · spread 0	16/17 · spread 0
NIST SP 800-63B (45k tokens), 13 q	11/13 · spread 0	7/13 · spread 2

The small corpus nearly saturates both. The real standard separates them, and the shape of the losses matters more than the totals.

GraphRAG loses specific values. Three of its misses have the answer sitting in the document while the reply talks around it — "verifiers should establish minimum length requirements … to enhance security" where the answer is 64, which never appears. That is its indexer working as designed: entity descriptions are summarised, and a summary abstracts the number away. It also asserted one confident wrong citation, OMB Memorandum M-04-04 where the document says Circular A-130.

LMM's two misses are refusals. Both are reworded questions whose answer is in the document and was not retrieved; both came back "I do not know" rather than approximated. Retrieval recall on synonyms is the open work item — and nothing was fabricated to cover it. Two instruments were built for exactly this class: an offline expansion channel, measured not to close it and since deleted, and the meaning channel that now ships switched on. See Honest limits.

Cost, on the 45k-token standard:

	LMM	GraphRAG
Ingestion	0 model calls (shallow — the mode a large document is for)	220 calls · 498,230 prompt tokens
Per question	~5 calls	2 calls · 10,132 prompt tokens

One correction this measurement forced on an earlier claim: "5× the model calls" was against vanilla RAG at one call per question. Against the real competitor at two, it is about 1.3×, and on the small corpus the per-question prompt token counts are nearly level (3,840 vs 3,306). Ingestion runs the other way in graph mode — 20× the calls — which is what per-fact extraction through the gate costs against batch extraction.

Two scorer bugs were found while running this, both against the baseline, both fixed. The abstention fallback was a Turkish-only phrase list, so GraphRAG's honest English refusals scored as fabrications (13/17 → 16/17 once fixed). And a token echoed from the question — "…required by the GDPR" — was read as a claimed designation, so refusing an absence question scored as inventing an answer. Handing points back to a competitor is not generosity; a benchmark that takes them by accident is not evidence.

Language independence, measured

No hand-written sentences, no word lists, no per-language rules anywhere in the pipeline. That used to be an argument about the code; every benchmark was Turkish, so nothing tested it.

corpus_en.txt and corpus_es.txt are the fictional corpus line for line — the same facts in the same order, the same 17 questions, the same gold keys translated. The invented names (zerbalit, vorlin, norgul, Nortlann) are identical in all three, so no answer in any language can come from a model's own knowledge. Only the language moves.

The same world, the same 17 questions	median of 3	samples
Turkish	17/17	17 · 17 · 17
English	16/17	15 · 16 · 16
Spanish	15/17	15 · 15 · 15

The measuring tool was the language-bound part. Scoring an abstention meant recognising "I don't know" — a phrase list, in one language. Against an English document every honest refusal would have scored as a fabrication. So the judgement moved inside: a turn now records whether it asserted anything, decided by re-extracting the claims from what was spoken. explain=True surfaces that stamp.

The two-point gap is one defect. Every remaining loss in English and Spanish is a refusal that keeps talking — "Vorlin is not a recognized substance", which the corpus flatly contradicts. Retrieval, derivation and the gates behave identically in three languages (the graph ingests 61–62 facts and derives 8 in each); the engine's stylistic habit of explaining its refusals varies, and this system's scoring is strict enough to charge it.

Cost, measured

Accuracy is only half of a comparison. The other half is what it costs, so the same two architectures were run again with a counter wrapped around the model clients — real usage tokens from the server, median of three repeats, both corpora. benchmarks/COST.md has the full tables; the short version is not flattering to us:

Per question, TR corpus	RAG	LMM deep=True
model calls	1.0	5.5
prompt tokens	593	3,944
wall clock	1.6 s	7.5 s
Ingestion (once)	0 API tokens (local embedding, ~0.2 s CPU)	42,449 + 1,174 tokens — or 0, with deep=False
Dependencies on disk	114 packages, 1.35 GB + 458 MB weights	1 package, 1.1 MB

On a hosted per-token engine, embedding RAG is still cheaper than LMM and there is no break-even — roughly 5-6x the calls and 6-7x the prompt tokens per question, and the gap widens with every question asked. Most of the remaining cost is the verification read-back: the gate re-extracting the claims out of a sentence before it may leave. That is fabrication-0 being paid for in tokens.

Where it goes now, and half of it you can stop paying today

That last sentence was true of the system that was measured, and that system no longer exists. The table above predates the shapes and the organs, and the cost harness's own list of headings was never extended — so every call the router and the organs made filed under other until 5 October 2026, by which time other held the largest number in the system. Re-measured on NIST SP 800-63B, 13 questions, a fresh process per arm (benchmarks/gate_arms.py):

bucket	calls
routing	21	which shape is this turn, which language, which two things
answer	19
extract	13
relation-check	12	the classifier readings
organ	8
judge	6
read-back	3	the check everyone optimises
widen	1
	83	6.4 a question

The read-back is three calls of eighty-three. It barely runs, because the two-tier gate admits a fully covered answer without asking the engine at all. What costs is routing — deciding what KIND of question this is, before any answer exists — and routing with extraction is 41% of a question.

Both re-discover something the caller usually knows. standalone=True says the question has no conversation on either side of it, which is true of every question an API receives, and shape= declares the kind. Same thirteen questions, one run each:

configuration	correct	calls	a question
as shipped	9 / 13	83	6.4
standalone=True	9 / 13	76	5.8
standalone=True, shape="material"	10 / 13	43	3.3

m.ask(question, standalone=True, shape="material")   # 6.4 calls -> 3.3

The eight questions the memory answers plainly fall from five or seven calls to two — one answer, one check. And no gate was traded for it, which is the only reason this is worth printing: read-back is 3 in both arms, so the exit gate ran exactly as often. What fell to zero was the classifier bucket, not the relation gate.

The extra correct answer is not claimed — the same arm has measured 9 and 10 on separate runs, so one run cannot tell a gain from noise. What is claimed is that nothing was lost for half the calls. Every question in this set is a document lookup, so material is the right shape for all of them; what a caller who declares the WRONG shape pays is not measured.

What the tokens buy is the other column of the table above — provenance, the refusal guarantee, symbolic multi-hop derivation, and a core with no dependencies. And on the local engine this project is actually built for, the dollar figure is zero and the cost is your own CPU seconds instead.

Zero-call answers — questions the graph answers with no model call at all — were 0 of 17 at every commit until lmm/lookup.py. The architecture always permitted a derived fact or a table row to be spoken as it stands; nothing in the code ever took that path, because respond always spent a call turning a held triple into a sentence and more calls reading that sentence back. lookup answers the question directly from the graph — the record, not a sentence — for the two shapes it can settle without guessing: a value named on a relation the graph has closed under derivation, or a field with one value left in it. Measured: 3 of 17 questions, in TR and in EN alike — 18% — now cost nothing at all, and they are exactly the multi-hop questions ("is vorlin a liquid", answered from the derived vorlin -[type]-> liquid). Everything the graph cannot settle safely still falls through to the paid path unchanged; see benchmarks/COST.md §3 for what that guarantee costs.

On a document whose facts are a TABLE, that floor is most of the questions. A spreadsheet's row label is usually several words — united states, new hampshire — and a scan of the question one word at a time could never name one, so those questions went to the engine. Naming a node by the question's own contiguous word runs settles them from the graph instead. Measured on the US Census spreadsheet fetch_documents.py pulls down, 10 questions, 3 samples, median (benchmarks/COST.md §8): 5 correct → 9 correct, 0 zero-call → 6, 6.1 calls per question → 2.7, with zero wrong answers in every arm. The three invented corpora do not move — their subjects are single invented words, so there is nothing multi-word to name — and they lose nothing either.

A row is only reachable by the name a question calls it. That sheet nests its rows by drawing the indentation with a character — .Alabama, .Puerto Rico, because a cell has no margin — and every nested row was invisible to the name anyone would use for it. The plain name is now an alias of the row: Unicode's category decides what is layout, and the cell keeps its own text, so nothing the document wrote is edited. Same protocol, 5 documents (benchmarks/COST.md §8.4): 48/50 → 49/50 correct, 15 → 16 answered with no model call, 0 wrong answers, the four other documents unmoved. The question that moved had been refused at every commit in this history and is now free.

A question with four right answers gets four right answers. That sheet's header occupies two rows, and reading only one of them overwrote three of its four population columns out of existence — silently. Reading the block (the sheet's own merges say where the header ends) keeps them, and then a question naming no year names four columns equally well; the answer path used to pick one and say nothing. It no longer chooses: a reading that names several records is answered with all of them, each under the full field name the SHEET wrote, which is where the year is. Four census questions that name a year were added first, because the old set never asked for a column the single-row reader destroys (benchmarks/COST.md §8.4.1, four arms, 5 documents × 3 samples): census 10/14 with 3 wrong answers → 14/14 with none, 6 → 10 of them free, field total 53/54 correct, 0 wrong, 19 zero-call, the four other documents unmoved. Both halves ship on; either can be switched off (LMM_XLSX_HEADER_BLOCK=0, LMM_LOOKUP_CANDIDATES=0).

And the same question over a memory that has not moved is free. Asking the same ten questions a second time costs 0.0 calls and 0.0 seconds on both field documents, every answer byte-identical. On the invented corpora it is 0.9 calls per question rather than 0.0, because two of the ten sit inside the research-offer flow and a replayed answer would leave the offer standing — so those two bypass the cache on purpose.

That floor is a property of ingestion, not only of lookup. The same corpus ingested by a weak local 3B model derived zero facts instead of Azure's 8, so lookup had nothing to answer from. It never guessed from the noisy graph it was given; it declined every time, correctly. The path itself spends no token once the graph exists — whether the graph closes at all still depends on what read it in.

Two defects stand between that engine and the floor, and neither is sufficient alone (benchmarks/COST.md §7.1, measured by replaying the local engine's cached readings through the real ingestion): its extractor names whole clauses as subjects, and — the larger one — its causal classifier answers YES for 41 of 61 sentences of this corpus, including plain is-a sentences the prompt's own examples cover, where gpt-4o-mini gets every one right. A sentence read as causal is written under #causes and never reaches the taxonomy, so the taxonomy is split across two predicates and no triangle can close, however clean the subjects are. A subject-normalization pass was built, measured against a PERFECT-normalization ceiling, found to buy nothing on either side of that split, and reverted. The remaining work is extraction quality on the local engine, not the wiring behind it.

Both of those were then attacked directly, and both attempts were reverted too (benchmarks/COST.md §7.1.1). llama.cpp can constrain its DECODER with a grammar, so a per-sentence GBNF was built in which a subject can only be a contiguous run of that sentence's own words and never the whole sentence — the wrong answer made unsayable rather than asked against — and beside it a sentence began to be written BOTH ways, as a causal edge and as a triple, so the taxonomy stopped being split. The engine answered by saying the longest thing still allowed (kelvit bir metaldir → kelvit bir). It is the first local run in which a predicate ever became transitive, and it derived nothing all the same, because the triangles that qualified it are triangles the corpus states outright. The 61 cached readings are kept beside the section so the next attempt starts from the measurement.

Architecture

                 ┌──────────── the GATE decides everything ────────────┐
 user/message →  extract (engine)  →  gate.admit  →  GRAPH (facts, sources,
                                                      trust, contradictions,
 document     →  learn_text ──────→  evidence index  transitive derivation)
 spreadsheet  →  learn_rows ──────→  graph directly (no model calls)
                                                      │
 question     →  lookup ──────────→  THE RECORD, spoken as it stands
                 (graph only)         no engine, no tokens, no wobble
                     │ not settleable
                     ↓
                 graph lookup + evidence retrieval →  answer (engine)
                                                      │
                 coverage gate → digit discipline → support check → verify
                 (an answer whose claims are not in the given evidence DROPS)

• The engine cannot write records and cannot speak unsupported facts. It produces candidates; the gate admits, the gate releases. Every other property here is downstream of that one sentence.
• Two speeds. Symbolic reasoning — derivation, causality, contradiction — runs in microseconds on the graph. The engine is only the language I/O.
• The graph answers first, where it can settle the question by itself. If the question NAMES a value on a relation the graph has closed under its own derivation ("is vorlin a liquid"), or names a field with one value in it, the answer is that record — returned before the extractor is reached, with no model call anywhere in the path and therefore nothing for a model to invent. Everything else falls through unchanged. It declines far more often than it answers, on purpose: nothing structural separates a question naming no field from one naming a field the graph has never heard of, so both are paid for. A node whose name is SEVERAL words — a spreadsheet's row label, a country, a spec table's field — is named by the question's own contiguous word runs, so united states is one name and not two misses.
• The same question over a memory that has not moved is not paid for twice. The answer is kept and re-spoken; teach the memory anything and it is dropped rather than repeated. What counts as "moved" is not a record count — it is every mutation the graph has, including a second source reinforcing a fact and a sleep() fading one, because a stale answer is worse than an expensive one. Memory(..., cache=False) turns it off.
• A document is two materials. Tables state their own structure and go straight to the graph with no model call; prose goes to the evidence index, where ranges and qualifications that don't fit a triple survive verbatim.
• Growth without retraining. Knowledge goes to the graph, never to weights. A periodic LoRA pass refreshes behaviour from gate-approved logs — never from the model's own raw output.

Full description: docs/architecture.md.

Install

pip install living-memory-model                      # core: one dependency, numpy
pip install 'living-memory-model[pdf]'               # pdfplumber + pypdf
pip install 'living-memory-model[xlsx]'              # pandas + openpyxl
pip install 'living-memory-model[docx]'              # python-docx
pip install 'living-memory-model[local]'             # torch + transformers + peft (local engine)
pip install 'living-memory-model[gguf]'              # llama.cpp, CPU
pip install 'living-memory-model[openai]'            # OpenAI, OpenRouter, Groq, vLLM, Ollama
pip install 'living-memory-model[azure]'             # Azure OpenAI

Extras are opt-in and lazy: nothing is imported until you hand LMM a file of that kind, and a missing reader reports the exact install line rather than a traceback.

One dependency, and it is named rather than hidden. The core used to be pure python with nothing under it. It now requires numpy, and what that buys is the meaning channel: a 30 MB static multilingual matrix travels in the wheel and is attached by default, so a paraphrased question is understood out of the box with no download, no account and no network. Measured, it retrieves better than the transformer it would otherwise take (MRR 0.88 against 0.71) and builds its index a hundred times faster. Memory(dense=False) runs without it; with numpy missing the channel is absent rather than broken.

The bundled matrix is derived from minishlab/potion-multilingual-128M (MIT, the model2vec method), reduced to 108k word-pieces and 128 dimensions; the segmentation is thirty lines of our own, so there is no tokenizer dependency either. See NOTICE.

Bring your own engine

LMM does not ship a model and does not want to pick one for you. It needs an engine to turn a retrieved record into a sentence, and any of these will do:

# nothing set — a local Qwen2.5-3B-Instruct, on your own machine
export LMM_MODEL_PATH=/path/to/any-instruct-model   # ...or any HF model you have

# llama.cpp / int4 — the fast option on cheap hardware
export LMM_BACKEND=gguf

# OpenAI
export LMM_BACKEND=openai
export OPENAI_API_KEY=sk-...
export LMM_MODEL=gpt-4o-mini

# OpenRouter, Groq, Together, a vLLM or Ollama server — same client, one URL
export LMM_BACKEND=openai
export OPENAI_BASE_URL=https://openrouter.ai/api/v1
export OPENAI_API_KEY=...
export LMM_MODEL=meta-llama/llama-3.3-70b-instruct

# Azure OpenAI — its own client, because a deployment name is Azure's shape
export LMM_BACKEND=azure

Nothing about the architecture changes with the engine, and that is the point. The graph, the gate and the derivation are plain python; they never call the engine and never learn which one is running. What the engine decides is how a record is worded — not whether it may be spoken. So a weaker engine costs you fluency, never provenance: an unsupported claim is dropped by the same rule whichever model produced it, and the calls that read, derive and verify are free on every one of these.

The gguf engine offloads to the accelerator when the installed llama.cpp build has one (it asks the library, not the platform); set LMM_N_GPU_LAYERS=0 to force the CPU, or a smaller number when the model does not fit in VRAM. Measured on an Apple M5, same call: 1.8-2.3 s on the CPU against 0.6-0.7 s offloaded (benchmarks/COST.md §7.2).

Using it

from lmm import Memory

m = Memory("mind.lmm")

m.learn("spec.pdf")                       # tables → graph, prose → evidence
m.learn("inspection.xlsx")                # rows → graph, milliseconds
m.learn("report.docx")
m.learn("The probe was built on Nortlann in 2031.")   # plain text works too

answer = m.ask("where was the probe built?", explain=True)
print(answer)                             # it IS the answer string
print(answer.abstained)                   # did this turn assert anything?
print(answer.sources)                     # the stamps it rests on

m.save()

learn() returns a small report (.facts, .tables, .adapter, .source).

A table inside a document is read row by row — a delimited row (| or tabs: markdown, ASCII tables) is whole on one line and its cells are stored under the names its header row gives them, so no window ever puts one row's value beside another row's label. A table of contents is read as what it is: a column of page numbers points at a place rather than stating a value, so it never enters the evidence — otherwise "what is this regulation's number?" is answerable with the page its section opens on, past every gate, because the page really is written beside the name.

The operator's knobs

Everything a RAG system prompt bundles into one block is a named parameter here, and each travels exactly where it belongs — none of them can reach the verifiers, because a judge whose thermostat the caller can turn is not a judge:

m = Memory("mind.lmm",
           persona="You are Ada, a warm onboarding coach. Ask before you advise.",
           style="One section per course: name, duration, audience, outcomes.",
           warmth=0.6,          # temperature of the VOICE surfaces only
           reply_tokens=1200,   # length cap of the voice surfaces only
           encoder=None,        # the bundled matrix; or any callable
           reranker="bundled",  # late interaction; or a callable, or None
           dense=True)          # the meaning channel, on by default

• persona is the voice. It rides in front of the phrasing prompts — chat, answer, refusal — and stops at the document's edge: the composer never sees it (measured: persona-flavoured rephrasing dies at the verbatim gate).
• style is the document's shape — the persona's mirror twin. It travels ONLY to the composer's request: sections, headings, ordering are the operator's; every fact beneath them still has to be attested by the material.
• warmth / reply_tokens tune temperature and length of the same voice surfaces. Unset, every surface keeps its measured default. The classifiers and the gate keep their pinned settings regardless.
• encoder is the meaning channel's vectors. Unset, the matrix that ships in the wheel. Any callable list[str] -> list[list[float]] replaces it — sentence-transformers, bge-m3, a vendor's embedding endpoint, an in-house service — and then nothing of ours is loaded.
• reranker is the ordering inside a proposal. "bundled" is the late interaction over the same matrix; any callable (query, texts) -> order replaces it (a cross-encoder, an LLM reranker); None turns it off.
• dense=False removes the channel entirely. The store then answers exactly as it did before the channel existed, and numpy is not needed.

# an operator who wants their own encoder, and no reordering
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("BAAI/bge-m3")
m = Memory("mind.lmm",
           encoder=lambda texts: model.encode(texts, normalize_embeddings=True).tolist(),
           reranker=None)

Two more are session attributes rather than constructor arguments, because they are about how hard a turn tries rather than how it sounds. Both default to 1, and both were measured down from higher numbers — the readings they bought stopped winning once retrieval improved (see Measured):

m.session.CANDIDATES = 3   # evidence subsets generated per question
m.session.VIEWS = 2        # views of the evidence a judge may buy

Serving it to more than one person

A memory that lives in one process is a library. lmm.serve is the same three verbs over HTTP, with one memory per user:

python -m lmm.serve --root ./stores --port 8000     # localhost by default
curl -s localhost:8000/learn -d '{"user":"ada","text":"Ada leads R&D."}'
curl -s localhost:8000/ask   -d '{"user":"ada","question":"who leads R&D?"}'
# {"answer": "...", "abstained": false, "sources": ["#doc:team.txt"], ...}

It is a transport, not a second brain: no gate, no rephrasing, no fallback text is added, so the abstention and the source stamps cross the wire intact and an HTTP caller can verify what an in-process caller can. Every user's store is chosen by their id and nothing else, no route reads across users, and an id that could climb a path is refused before it becomes a filename. Each memory holds its own lock — one writer at a time per user, different users in parallel. Set LMM_TOKEN to require a bearer token.

TypeScript callers get the same object rather than a bare string (sdk/typescript, no dependencies):

const m = new LMM({ base: "http://localhost:8000", user: "ada" });
const a = await m.ask("who leads R&D?");
if (a.abstained) console.log("memory does not hold this");
else console.log(a.answer, a.sources);

Inside a LangChain application

from lmm.adapters.langchain import retriever, tool

chain_retriever = retriever(m)    # passages, each stamped with its document
agent_tool = tool(m)              # an AUDITED answer, with its sources

The retriever is ordinary retrieval — the gate is not in play, and what the caller's own model then says about those passages carries the caller's own guarantees. The tool is the one worth having: the agent receives the answer that passed the gate, and receives the refusal as itself, because an agent handed an empty string writes its own answer over the silence. LangChain is not a dependency; without langchain_core installed the same calls return duck-typed equivalents.

A conversation that may not teach

m.session.respond(msg, teach=False) is the consultation surface — a customer-facing bot that listens, remembers the conversation, and delivers, without ever letting a visitor write into the operator's memory:

• a statement ("we are a bank, team of ten") is CONTEXT — it routes to the chat voice and its terms accumulate in the consultation's brief;
• a question searches with the whole consultation riding along, so "the best three-hour material" still knows the topic named two turns earlier;
• a delivery request ("you decide, put a programme together") goes straight to the composer, which returns a source-stamped catalogue built from the brief — and echoing the user's own words back is conversation, not assertion, so the gate lets a consultant sound like one.

Pass on_line= to respond or compose and the draft STREAMS: each line is judged by the composer's gate the moment its newline arrives and handed to your callback while the engine is still writing. A line the gate refuses is never seen — an invented number does not become printable by arriving early. On a backend that cannot stream, the same code degrades to batch by itself.

The whole surface

Five methods cover ingesting and asking. Everything else this README claims — contradiction, causality, the trust ladder, the memory dynamics — lives on m.session, which is public and is the same object Memory is built on. It is listed here because a feature with no documented entry point is not a feature.

m.learn(what)	teach it a file or a string	no engine for tables
m.ask(q, explain=True)	answer, with .sources, .abstained and .route (which organs the turn passed through) — cannot write memory	engine
m.about(label)	the records held on a concept	no engine
m.facts	how many records exist	no engine
m.save(path)	graph and evidence, both	no engine
m.compose(brief, topics=, on_line=)	a structured draft from the evidence — per-topic gathering, gated line by line, streamed to on_line as lines survive, returned with its sources	engine
m.where(term)	which documents mention this — names and counts, the census	no engine
m.themes()	which documents belong together, and on what entities — community detection over the store's own graph	no engine
m.distil(text, source=, speaker=)	write the EVENTS a passage reports into the graph — the engine lists them, the passage's own words admit them, each survivor is one gated dated record	one call per passage
m.bridge()	teach the store, once, what words readers ask its fields with — afterwards those questions are answered by the record itself	one call per field, once
m.session.respond(msg, teach=False, on_line=)	a conversational turn — teach=False is the consultation surface (context, brief, delivery; cannot write memory)	engine
m.session.learn_rows(rows)	[{column: value}] straight to the graph	no engine
m.session.learn_cause(a, b)	record that a causes b	no engine
m.session.root_causes(x)	walk the causal chain back	no engine
m.session.causes_of(x) / .effects_of(x)	one step either way	no engine
m.session.tension()	the contradictions it is holding	no engine
m.session.curiosity()	what it has been asked and cannot answer	no engine
m.session.sleep()	fade, reinforce, settle episodic into semantic	no engine
m.session.verdict(old, new)	arbitrate two rival values	engine

A record is a core.memory.Record: .subject, .predicate, .value are concept keys, not strings, because one spelling can be two entities and one entity can carry labels in several languages — read the labels through m.session.memory.identities. .sources and .trust are the provenance the gate reads on the way out.

The column that says no engine is not a footnote. Those calls are plain python over a dict — microseconds, no network, no key — which is why a spreadsheet loads in milliseconds and why the reasoning half of this system costs nothing to run.

See examples/ — including one that runs with no engine at all.

What 0.7 changed — conversation memory learns what "living" means

A told fact can END, a question is asked ON a day, a thing HAS a kind, and a small telling can be read whole. Six mechanisms, all generic (no word of any language in code), each behind a failing test first — measured on one fixed 30-question chat-memory slice throughout, with the document exam re-run after as the control.

change	measurement
the graph count asks the engine which facts still stand	"how many subscriptions" flipped between a fabricated 3 and a short 1; now the right two entities, cancelled one pruned, stamped
the plan knows the asking day (session.asked_at, the now primitive)	"how many days ago…" abstention → the exact answer
a distilled thing carries its KIND — an index key, never a claim	"I'm also getting Architectural Digest" reachable by "magazine subscriptions"
a list is not a name; membership by name or kind, never by story	a projects count that swallowed tools and clauses: 8 → 2
a stated tally outranks an enumeration; one breath is not a series	four stated-tally answers ("20 playlists", "32 species"…) restored
a small dated telling is read whole when every organ dies (its own gates)	"which did I use most recently", "what are the two hobbies": refusal → stamped answers
the slice, before → after	43% → 57-60% · wrong claims 7 → 5-6 · document exam 30/31, unchanged · invariants 211 → 218

A seventh mechanism was built, measured and reverted before release: handing these turns to the factual chain contradicted a measured law (the chain's 23 confident wrong claims were measured on chat). The whole-telling organ is the corrected design — a new reading with its own gates, not a loosened chain. Undated stores — every document corpus — build their blocks byte for byte as before, and the tests assert it.

What 0.6 changed, and what it cost

Every line here was measured, and three of the changes are RETRACTIONS of insurance this project had been paying for.

change	measurement
the meaning channel ships and is on	answering line reaches the engine 32% → 79%
it sits at the DOCUMENT layer, not the line	store retrieval 92% → 99% (103 known-item queries)
late-interaction reordering	worth more the vaguer the question: 6 terms removed, MRR 0.707 → 0.807
the candidate ladder, 3 → 1	eleven field questions: 139 → 104 calls, correctness 4/11 → 5/11; the narrow candidates won nothing
a judge's second view, 2 → 1	the second view confirmed nothing, on either judge; −8 calls in 104
the derived index is written beside the store	first question 10.6 s → 0.13 s
a deterministic question is asked once	one refusal turn: −2 calls, −11k tokens
the offline expansion channel	deleted, −607 lines

The three retractions share a shape. Each was a hedge against weak retrieval — generate three readings and pick, judge against two views, index the questions a line might answer — and each stopped earning the moment retrieval got better. None was deleted outright: CANDIDATES and VIEWS are numbers, and an operator whose corpus defeats the reordering raises them.

---

What the gate is for — corrected by a field measurement

A team integrating this library published a correction to their own report, and it corrects the argument this README had been making. Their first comparison put the whole document in the prompt — no retrieval at all — which made the gate look like it was buying two fewer invented answers.

Run against the right control (the same retrieval LMM uses, one engine call after it instead of the gate), across 44 supplier/requirement cells on two real tender documents:

	tokens	seconds	false-covered
document in the prompt	292,749	125	5
the same retrieval, one call, no gate	28,083	88	3
LMM	313,173	491	2

Retrieval removes two invented coverages; the gate on top of those same lines removes one more — inside a variance of about two cells. On that corpus the gate's contribution to non-fabrication is at the noise floor, because once retrieval is good the fabrication it exists to stop was mostly not happening. Naive RAG over the same retriever was 11× cheaper in tokens and 5.6× faster, and anyone choosing between them deserves to know that.

What the gate does buy showed up in the same run, and neither side had measured it. Counting what each system was willing to say:

	"partly"	"fully"	"not covered"
same retrieval, no gate	31	7	6
document in the prompt	14	29	0
LMM	0	34	10

Unverified answering hedges: 31 of 44 cells came back "partly", and the whole-document variant never once said "not covered". LMM commits, 34 to 10, and 8 of those 10 refusals are right. A matrix where 31 rows of 44 say "partly" tells its reader nothing.

So the claim is not "the gate stops the model inventing things". With good retrieval, it mostly was not. The claim is: the gate stops a system from refusing to answer while appearing to. A system that never commits is never wrong and never useful.

(The same run reports zero fabrications in 44 cells. Both of its wrong answers were relevance failures — see Honest limits.)

Honest limits

Kept current, and deliberately specific.

• Cross-session recall over chatty dialogue is partly closed, and the remainder is named. 0.7's kind keys and whole-telling reading answer the cases that were the face of this wall ("getting Architectural Digest" now counts as a magazine subscription; "what are the two hobbies" answers, stamped). Still open, each a missing derivation rather than a mystery: summing durations across records ("how many years in total"), ordering THREE events (the order organ compares two), reflecting a stored preference inside a recommendation, and a ±1-day anchor when the event's date is written in the text but the envelope's stamp wins. One more is variance, not architecture: extraction at temperature 0 is not byte-stable across processes, and a 30-question score moves ±2 with it.

• On a fixed 30-question LongMemEval slice (5 per category, gpt-4o-mini as the engine), 0.7 measured 57-60%, from 43% at the start of the cycle. That figure is no longer reproducible and should not be relied on: its harness and its question list lived outside this repository and are gone, so which data variant it used is unknown. 0.12.1 re-measured with a harness that is now IN the repo (benchmarks/longmemeval.py, a deterministic slice) and scores 43% on the oracle variant and 10% on longmemeval_s, judged by an engine. The 33-point gap between those two is distractor sessions, and it is the honest shape of this weakness. 0.13.0 measures 13/30 on longmemeval_s (quoted=True, the shape read rather than declared), where 0.12.1 measured 7/30 in the nearest comparable configuration; the per-question files are in benchmarks/longmemeval_runs/. Two things limit what that figure says: thirty of five hundred questions, so one question is more than three points; and the judge is the configured engine, gpt-4o-mini, the same model that wrote the answers. Published chat-memory figures (Zep's current claim is ~71% under a GPT-4o judge) use different engines and protocols. That benchmark is a long-personal-conversation exam, not a document exam; the same code answers 40/40 on the field set above and its document exam did not move during any of this. Both numbers are true and neither is the whole picture. What each release moved is named in the changelog; what it did not is named here — including that roughly half of what we still miss, the bare engine handed the same lines also misses, and misses confidently where this memory abstains.

• A count verifies that its items exist; membership is checked only where the set is a stretch of time. Asked how many doctor's appointments the user went to in March, the counting organ named three doctors where the answer is two: every name was written in the evidence, so every name passed, and one appointment was on April 1st. Fixed for dates — each line is given the date its own sentence states rather than the date the message carries, the question's window is read as two dates, and the library compares them. That question answers two now. Not fixed for anything else: a doctor the user was only CONSIDERING is excluded by no word of the question and by no date. A word-level repair was built, measured and refused, because the same arithmetic that drops the April appointment drops a true project from the invariant guarding this organ. Every organ that verifies items against lines inherits the remainder.

• Relevance is measured now, and it is 8 of 9 either way. The relation check keeps 8 of 9 relevant answers and stops 8 of 9 irrelevant ones on the NIST set, with the negatives built rather than written: each question paired with a true, sourced sentence from the same document about something else (benchmarks/relevance.py). Nine questions, so one question is eleven points. Both errors are the same question, and it is the paraphrase one — this document answers "what is the shortest password" under memorized secret, and the gate cannot see that the answering line answers it. That gap shows in retrieval, in the rewording pass, and here.

• Relevance is the dominant failure mode, and the gate does not touch it. The gate guarantees non-fabrication, not perfect relevance — and a field integration measured that this is not one limit among several but the one: zero fabrications in 44 cells, both wrong answers relevance failures. One is instructive: every content word verbatim in the document, a real source stamp, a true sentence — the requirement asked for a simulation feature in the product, the document described unit testing during the project. The gate had nothing to object to, because nothing in it was false. "The memory has something to say about X" is not "the supplier offers X", and nothing here judges that distance.

• Spec lines reachable only through very common words ("how many inches is the screen", where the line reads Screen 15.6" LCD and never says inches) are a lexical-retrieval ceiling. One such case is rescued by an append-only mechanism; the class is not closed.

• A question that names PART of a column's name and nothing that tells the columns apart ("the 2021 population estimate", where the sheet's column is Population Estimate (as of July 1) 2021) is not answered from the graph at all — it is one reading of several and the widest reading holds one record, which this path does not speak. It goes to the engine like any other question and is answered there, at the engine's price. What is guaranteed is that the three columns the question ruled out are never spoken over it (benchmarks/COST.md §8.4.1).

• The offline expansion channel was deleted, and the case it was built for is still open. It generated the questions each line answers and indexed them, keeping a query when it reached its own line better than any other. Measured on NIST SP 800-63B that filter is inverted: the line reads "Memorized secrets SHALL be at least 8 characters in length" and a reader asks "what is the shortest password", which shares not one distinguishing word with it — so a real rephrasing never reaches its own line and the filter kept only the queries that COPIED it. Handed perfect queries by hand, it kept zero. A switch nobody should turn on is a liability, so it is gone. The meaning channel that replaced it closes much of the class (the answering line reaches the engine 32% -> 79% of the time on a 115k-line corpus) but not this example: measured again after, "what is the shortest password" still does not reach that line, while "minimum length for a memorised secret" reaches it first.

• Two benchmark questions are stable failures and are named rather than hidden: one whose answer sits in a table row sharing a single stem with the question, and one needing a heading plus a line eight sentences below it in one window.

• The engine sometimes appends an explanation to a refusal that the memory contradicts. This is detected — the turn is scored as having asserted something, and loses the point — but not prevented. It is the whole of the remaining English/Spanish gap. (The other source of appended text, a second model call after the gate, is gone: the hedge is now the stored source stamp in brackets, (~ #pdf:manual.pdf), and nothing is appended to a refusal at all.)

• LMM does not OCR. A PDF without a text layer — what an office scanner produces — teaches it nothing, and learn() now says so instead of reporting a cheerful zero. Run OCR first and learn the result.

• There is no HTML reader. An .html file is read as plain text and its markup lands in the evidence index; learn() warns, with the share of characters that were tags. Convert to text first.

• Table columns are still flattened by the PDF reader: a cell can reach the index next to a neighbouring column's value, and no gate can reject that, because the claim really is in the evidence. Rows are repaired; columns are not.

• A spreadsheet whose header spans more than one row loses the unnamed columns.

• An engine call is bounded by LMM_TIMEOUT (90 s by default; 0 removes the limit) on the API backend, which raises EngineTimeout rather than waiting out a quota. The local engine's generation loop is not interruptible, so the budget does not cover it.

• The coverage gate cannot distinguish negation affixes at word level (language lists are forbidden by design); the engine-level support check covers most of that class.

• Pre-1.0. The API surface may still move.

Contributing

Please read CONTRIBUTING.md — it leads with the two rules that make this project what it is: no document-specific constants and no hand-written language rules. Both are stated with what enforcing them cost.

python3.11 tests/test_core.py     # 266 invariants · no model · no GPU · no network

Every test is a pathology that was measured on this code, written back as an assertion so it cannot return.

Citing

The design, the measurements and their corrections are written up in a report with a permanent DOI:

KANAR, Ruzgar Ersin (2026). Living Memory Model: A Memory Layer That Does Not Say What Its Evidence Does Not Support. Zenodo. https://doi.org/10.5281/zenodo.22999474

That DOI always resolves to the latest version of the report. GitHub's "Cite this repository" button reads the same entry from CITATION.cff.

Licence and attribution

Apache 2.0 — see LICENSE and NOTICE.

LMM is not a language model trained from scratch. It is a memory and reasoning layer on top of one. The default engine is Qwen2.5-3B-Instruct (Apache 2.0, Alibaba Cloud), downloaded by the user from its own source. The local engine can also load a LoRA adapter the user supplies; none ships with the package and nothing in this repository trains one. What is original here is the layer — the graph, the gate, the evidence index and the adapters. The base model is credited, never hidden.