nanomem

An embedded store for facts that change.

A fact your application remembers is not a document. It gets corrected — people move, change jobs, switch phone numbers. A vector store keeps both statements and returns whichever one is worded closer to the question, which is how an assistant ends up confidently repeating an address you left two years ago. nanomem keeps the chain and knows which end of it is current.

Evidence — how it compares to FAISS, sqlite-vec and Chroma (including where it loses), what happens when the process is killed mid-write, and the pytest command that re-runs most of those claims on the copy you just installed.

pip install nanomem

import time
from nanomem import Vault

DAY, now = 86400, time.time()
job = {"entity": "employer"}

v = Vault("memory.dat")
v.add("I work at Acme Corp.",                    metadata=job, timestamp=now - 300*DAY)
v.add("I moved jobs, I now work at Initech.",    metadata=job, timestamp=now - 155*DAY)
v.add("I switched again, I work at Globex now.", metadata=job, timestamp=now - 10*DAY)

print(v.search("where do I work")[0]["text"])
# I switched again, I work at Globex now.

for r in v.history("where do I work"):
    print(r["revision"], r["superseded"], r["text"])
# 1 True I work at Acme Corp.
# 2 True I moved jobs, I now work at Initech.
# 3 False I switched again, I work at Globex now.

print(v.search("where do I work", as_of=now - 200*DAY)[0]["text"])
# I work at Acme Corp.

f = v.volatility()[0]
print(f["entity"], f["n_revisions"], round(f["median_interval"]/DAY))
# employer 3 145

One file on disk. One runtime dependency (numpy). No server, no daemon, no index to rebuild. Search is exact — a full cosine scan, not an approximate index — so recall is 100% by construction and every interesting question is about time rather than ranking.

Tell it what an attribute is

metadata={"entity": "employer"} is doing real work above, and it is worth a paragraph because little else here matters as much.

Name the attribute and nanomem knows those three statements are one fact, so it keeps them as a chain. Leave it out and a lexical tagger guesses from the text — measured recall 70 of 100 on plainly-worded chains and 0 of 100 on narrative phrasing. In the example above it tags "I moved jobs, I now work at Initech." as location rather than career, because "moved" outweighs "work at", and the chain silently splits in two.

So if your application has attributes of its own, declare them. Everything nanomem does that a vector store does not rests on knowing which statements are about the same thing — and you know that, while the tagger is guessing.

Use it from Claude, Cursor or Zed (MCP)

pip install nanomem

Then add this to your MCP client's config — claude_desktop_config.json on macOS lives at ~/Library/Application Support/Claude/:

{
  "mcpServers": {
    "nanomem": {
      "command": "python3",
      "args": ["-m", "nanomem.mcp", "--vault", "/absolute/path/to/memory.dat"]
    }
  }
}

Restart the client. It gains seven tools: nanomem_add, nanomem_search, nanomem_history, nanomem_as_of, nanomem_changes, nanomem_volatility, nanomem_stats — so the assistant can ask what a fact USED to be, what the memory believed at a past moment, what changed last week, and which of its own beliefs have gone stale.

The vault is an ordinary file. Point the CLI or a Python script at the same path to read what the assistant wrote — a write is on disk before its reply is sent, so another process sees it immediately and stopping the server cannot lose it.

Through 0.7.17 that was not true: the server flushed only on a clean exit, and an MCP client stops its servers with SIGTERM. Twenty nanomem_add calls, each answered "Stored …", then SIGTERM, left zero rows in the vault. If you ran an earlier version, anything the assistant "remembered" in a session that was not closed cleanly was never written.

Package 0.7.19 · engine 3.4.6 · container format 3 · arena cache format 3.

Licence: AGPL-3.0-or-later, or a commercial licence. Free for personal, academic and open-source use, and for running internally on your own machines. If you offer nanomem to users over a network, AGPL section 13 requires you to offer them your source — see COMMERCIAL-LICENSE.md for the alternative. Up to 0.6.0 the wheel metadata said Apache-2.0 while the LICENSE file said All Rights Reserved; that contradiction is resolved here and the superseded terms are kept in LICENSE.preview-v1.0.md.

Run the demo

cd nanomem_standalone
python3 demo.py

It stores a few facts, updates one of them to show revision handling, searches with citations, and prints the vault's real stats() — document count, file size, the measured active_heap_ram_kb, and whether the file is encrypted (by default it is not).

Run the tests

python3 -m pytest -q

697 tests, no network needed.

There is no test_security.py. Earlier versions of this README told you to run one to "prove that zero plaintext exists on disk"; that file never existed, and the claim was wrong anyway — a vault is a plaintext file unless you give it a passphrase. To check for yourself:

python3 -m nanomem.cli init demo.dat
python3 -m nanomem.cli add "the office wifi password is hunter2" --vault demo.dat
strings demo.dat | grep hunter2          # plaintext vault: it is there
python3 -m nanomem.cli rekey --vault demo.dat --new-password-stdin
strings demo.dat | grep hunter2          # password mode: it is not

Use it from your own script

from nanomem import Vault

with Vault("my_knowledge.dat") as vault:            # plaintext by default
    doc_id = vault.add("Server backup runs daily at 02:00 UTC")
    vault.add("The staging database is on port 5433")

    hits = vault.search("When does the backup run?")
    print(hits[0]["text"], hits[0]["cosine"])

add() returns the document id, so you can get, update or delete by it later. For an encrypted vault, pass password="…" (or set NANOMEM_PASSWORD).

Embeddings are yours to choose. The default is nomic-embed-text on a local Ollama-compatible daemon, but any model of any width works — the width is probed from the model itself, and the vault is sized from what it returns:

Vault("m.dat", embed_model="all-minilm")                    # 384-d
Vault("m.dat", embed_model="mxbai-embed-large")             # 1024-d
Vault("m.dat", embed_model="text-embedding-3-small",
      base_url="https://api.openai.com/v1/embeddings")      # 1536-d
Vault("m.dat", embedder=MyOwnEmbedder())                    # anything with
                                                            # .embed/.embed_batch/.dim

Point it anywhere with base_url= or NANOMEM_EMBED_URL. Pass dim= to EmbeddingProvider to skip the probe entirely (air-gapped installs). An existing vault's width always wins over the one you request, so you cannot silently corrupt a vault by naming a different model later — you get a warning and the file keeps its own width. That handle is then unusable: your encoder is still the model you named, so every call raises until you reopen with one of the file's width. Nothing is lost, but it does not quietly carry on.

Without a daemon there is a fallback, and it is worth knowing what it is: a deterministic lexical encoder at 768-d — words and character trigrams hashed, no semantics. It scores "the server is up" against "the server is down" at 0.783, and "I drive a car" against "I own an automobile" at 0.286. Opposites look identical, synonyms look unrelated. It keeps the pipeline running offline and is fine for a smoke test; it is not a substitute for an embedding model, and there are no bundled neural weights.

What makes it different from a vector store

nanomem is an append-only log, so it keeps every value a fact has ever had, not just the current one. That makes four questions answerable that a vector index cannot represent, because none of them keeps the history to answer from.

with Vault("memory.dat") as v:
    v.history("where do I work")          # every value, oldest first, current last
    v.search("where do I work",
             as_of=1735689600.0)          # the answer as the memory stood back then
    v.changes(since=1735689600.0)         # what was written in a window, no query
    v.volatility()                        # how often each fact actually changes

as_of and since are unix timestamps in Python; the CLI below takes YYYY-MM-DD as well. history() returns each value with its timestamp and a superseded flag; the last entry is the current one. A fact that never changed returns one entry, which is an answer, not an empty result.

volatility() is the one to look at. It reads the revision log and reports, per fact, how many times it has been restated, the typical interval between changes, and how long the current value has stood unconfirmed — so an agent can work out which of its own beliefs have gone stale and ask again:

fact             restated   changes every   last confirmed
mobile_phone            4          195 d           2237 d   ← ask again
employer                3          807 d             47 d

No model and no query: differenced timestamps off two resident columns, off the search path. staleness() will turn that into a probability, but it returns None unless you pass assume_memoryless=True — the per-fact rate only beat a single corpus-wide rate on data generated to match its own assumption, so it is not on by default (evidence/staleness_calibration.json).

What bounds this. Two records are only treated as one fact when nanomem can tell they are the same attribute, and that needs either a vocabulary or the query — five vocabulary-free signals were measured and all are at chance against sibling attributes like home-vs-office address (evidence/grouping_signal_results.json). With canonical phrasing the tagger groups 70 of 100 chains; with narrative phrasing ("ported the line over the weekend, reach me on …") it groups 0 of 100. If your application knows its own attributes, declare them — pass metadata={"entity": "employer"} on write — and set group_floor_sim=0.45, worth +19.0 points of top-1 on exactly the phrasing the tagger struggles with (floor_retune_results.json). Leave both alone if you are relying on the tagger; the same setting costs 13.9 points there.

Declare it wherever you write from. metadata={"entity": ...} in Python, entity on the MCP nanomem_add tool, --entity on nanomem add. Until 0.7.12 the last two did not exist, so the two surfaces most callers integrate through were locked onto the tagger with nothing saying so.

What it costs when you cannot declare one. On a group the tagger inferred, search and history can disagree about which value is current: history resolves the tagged chain and reads the revision counter, while search also applies a relevance floor that can drop the newest value when it is phrased further from the question than an older one. Exempting the newest member of an inferred group was measured and rejected — it costs -19.4 points on the 3-persona chat set, and a sweep from 0.40 to 0.80 found no threshold that bought the one without paying the other (evidence/floor_current_value_results.json). So this is a deliberate trade, not an oversight.

0.7.12 said here that history is authoritative and should be believed over search. That was wrong and the claim was withdrawn: history could itself return a shorter chain than existed, in the worst case one entry flagged superseded=False, which this library defines as "this fact never changed". That truncation is fixed in 0.7.14 and 0.7.15 — a declared chain now returns every revision it holds, and changes() agrees with it.

What remains is the disagreement itself. Until 0.7.16 this section said "declare the entity and the disagreement goes away", on the strength of three chains scoring 3/3 (evidence/entity_declaration_results.json). A third independent black-box review found a declared chain where it does not, and the claim was wrong in a way the measurement could not see: in all three of those chains every revision RESTATES the attribute ("I now work at Initech"), which keeps the chain's members close together. When revision 1 names the attribute and later revisions refer to it implicitly -- "My desk is on the third floor of Kestrel House" then "Now parked in the Maple Wharf building" -- the members spread apart, and the lift that promotes the current value was capped below what that spread needs. search[0] returned a desk location two moves old while history() returned the right one, with the entity declared.

0.7.16 raises that cap on a sweep over eight arms and two embedders (evidence/revision_lead_cap_results.json). Measured over twelve chains of the phrasing that broke it, each alone in its vault, in two timestamp regimes:

and left to the tagger stays 2/3, unchanged and still measured on three chains only (evidence/entity_declaration_results.json).

Declaring the entity is what makes the chain resolvable, and on a real embedding model it is what makes search and history agree on these sets. It is not a guarantee. On the twelve offline-encoder chains measured here, every remaining failure has the same cause: that encoder is lexical, the newest revision lands so far from the question that the relevance screen drops it from the result entirely, and no ranking boost can promote a record that was never returned (6 of 6 failures, revision_lead_cap_results.json arm H).

That is the cause on THIS corpus, not the only one there is. An independent reviewer, building chains of the same described shape, measured failures of a second kind: the newest revision IS returned, the boost is applied and saturates at max_boost, and the lift needed to lead the chain (~0.68) still exceeds the 0.60 cap. Those would be fixed by a larger cap; the ones measured here would not, which is why the published sweep shows this arm flat from 0.60 through 1.50. Two corpora of the same shape can differ this much on a lexical encoder, so treat the 75.0% as what it is — a measurement of twelve chains, not a property of the fallback. tests/test_revision_lead_cap.py fails if the cap is ever reached on the sets it checks. changes() is the unfiltered view to cross-check against.

Building RAG on it

Retrieval is the core; generation is optional and uses no extra dependency (stdlib urllib to whatever endpoint you point it at).

v.search("what port does staging use?", top_k=3)     # retrieval only
v.ask("What port does staging use?", llm="llama3.2") # -> {question, answer, citations}
v.ingest_file("runbook.md")                          # chunk + store a document

nanomem proxy additionally serves an OpenAI-compatible /v1/chat/completions that injects memory, so an existing app can point at nanomem instead of its provider and gain memory without code changes.

Providers. Pass base_url= and api_key= to ask()/chat(), or set them on the Vault. The base URL decides the request shape:

Embeddings come from NANOMEM_EMBED_URL (default nomic-embed-text on Ollama). Routing is decided by the PARSED PORT, not by a substring of the URL — before 0.6.1 a host named web8000.internal was misrouted and Anthropic was POSTed to /v1/chat/completions, which 404s.

CLI

python3 -m nanomem.cli init company.dat
python3 -m nanomem.cli add "DB port is 5433" --vault company.dat
python3 -m nanomem.cli search "DB port" -k 3 --vault company.dat
python3 -m nanomem.cli history "DB port" --vault company.dat
python3 -m nanomem.cli search "DB port" --as-of 2026-03-01 --vault company.dat
python3 -m nanomem.cli changes --since 2026-01-01 --vault company.dat
python3 -m nanomem.cli stats --vault company.dat
python3 -m nanomem.cli rekey --vault company.dat --new-password-stdin

add prints the document id and its own elapsed time.

What it does, measured

Search is an exact linear scan. For the text it scans, it returns what an exhaustive fp32 cosine scan returns — 0 of 120 top-4 order differences on a 1,190-document corpus (evidence/exactness_v3r2.json) — and its latency therefore grows with the corpus.

Which text it scans depends on decompose, which defaults to True. A multi-clause question is split and each sub-query is scanned exhaustively, then the best hits are interleaved — so the result is the exhaustive answer to each clause, not the exhaustive top-k of the whole sentence, and for a multi-clause question the two differ. That is the point of decomposition: a two-part question gets both parts answered. Pass decompose=False when you want the whole string treated as one query and the exactness claim above to apply end to end.

evidence/scale_results_v3r4.json, headtohead_v3.json, rss_v3r4.json. Measured on Apple M4 Pro / Python 3.12, nomic-embed-text 768-d, random insertion order, 500 held-out questions, ingest → close → re-open → search. (Every evidence/… path in these documents is relative to the repository root, not to this folder; the JSONs and the scripts that wrote them live there.)

The "same as exhaustive" columns are recall. Ordering is identical too at 1,190 documents (0 of 120 top-4 order differences); at 10,000 and 71,433 a few tie-breaks differ — 2 of 500 and 5 of 500 questions, largest cosine gap 2.2e-05 — because the on-disk vectors are fp16 (evidence/verify_round3_v3r3.json).

The scores are exact in the same sense and not a bit further: two float32 matmuls of different shapes reduce in different orders, so the score attached to a hit can differ in its last bit or two between one BLAS and another — measured 2.98e-08, two ulps, between Apple Accelerate and OpenBLAS on the same query. The answer does not move with it. Over 4,000 documents at 128 and 768 dimensions the worst |fp32 − fp64| error is 2.01e-07 while the smallest gap between rank 4 and rank 5 is 4.46e-06 — twenty times larger — and 0 of 150 queries were undecided at k = 1, 4 or 10 (evidence/screen_exactness_results.json). So "returns what an exhaustive scan returns" is a claim about which documents come back and in what order — except among documents the ranking genuinely cannot separate, where the corpus contains no tie-break to be faithful to. It is not a claim about the bit pattern of the float beside them, and it never could have been.

Resident memory is roughly 8–9 KB per document. Earlier documentation claimed a "< 500 KB RAM" or "160 KB active heap" footprint; those were constants printed by stats(), not measurements, and they are gone.

Writes are flat as the vault grows: 19.81 µs mean over the first 500 of 4,000 adds, 19.32 µs over the last 500, excluding embedding (evidence/headtohead_v3.json).

Deleting one record is a full atomic rewrite — 7.5 ms at 1,000 records, 70.5 ms at 10,000 (evidence/rewrite_cost_v3r4.json). There are no tombstones in 3.0.

Documentation

• USER_MANUAL.md — hub, and the full measured-performance tables
• USER_MANUAL_PERSONAL.md — chat, profiles, privacy
• USER_MANUAL_DEVELOPER.md — SDK, CLI, concurrency, migration
• SERVICES_AND_API_SPECIFICATION.md — REST endpoints and limits
• MULTIHOP_REASONING_AND_TOPOLOGY_GUIDE.md — the bridge, and what did not work
• CHANGELOG.md — breaking changes from 0.1.x

Read nanomem.THREAT_MODEL before relying on the optional password mode. It is scrypt + a SHAKE256 keystream + an HMAC-SHA256 tag, built from the Python standard library. It is not AES, not "256-bit encryption", and it has not been audited.