Cuba-Memorys

Long-term memory for AI coding agents. An MCP server that gives your agent a knowledge graph it can search, reason over, and be corrected by — so it stops forgetting your codebase between sessions.

Written in Rust. Backed by PostgreSQL + pgvector. 28 MCP tools (29 with CUBA_DOCS=1), 22 CLI commands, and every number below measured on a benchmark that — as of v0.12 — actually measures what it claims to. (The previous one did not. See Measured.)

Install

pip install cuba-memorys        # or: npm install -g cuba-memorys
claude mcp add cuba-memorys -- cuba-memorys

That is the whole setup. On first run it provisions a PostgreSQL 18 + pgvector container via Docker and initializes the schema. Docker must be running.

{
  "mcpServers": {
    "cuba-memorys": {
      "command": "cuba-memorys"
    }
  }
}

{
  "mcpServers": {
    "cuba-memorys": {
      "command": "cuba-memorys",
      "env": { "DATABASE_URL": "postgresql://user:pass@localhost:5432/brain" }
    }
  }
}

cuba-memorys serve                      # 127.0.0.1:8787 by default
cuba-memorys serve 127.0.0.1:9000       # or pick the address

{
  "mcpServers": {
    "cuba-memorys": {
      "type": "http",
      "url": "http://127.0.0.1:8787/mcp",
      "headers": { "Mcp-Client-Id": "editor-window-1" }
    }
  }
}

cuba-memorys models all          # embeddings + NLI + reranker + runtime
cuba-memorys models embed        # just the embeddings model (~113 MB)
cuba-memorys models all --gpu    # GPU runtime, if you have one
cuba-memorys doctor              # confirms what loaded

What it actually does

Most memory servers are a key-value store with an embedding bolted on. This one models four kinds of memory, because the psychology literature says they are four different things and they decay differently:

Procedural memory is a separate table rather than a ninth observation type for a specific reason: ACT-R separates declarative memory (reinforced by access) from procedural (reinforced by success). As an observation, a recipe consulted constantly because it keeps failing would climb in importance. It is ranked by Wilson lower bound, so 1/1 successes scores 0.21 and 47/50 scores 0.84 — a lucky first try does not outrank a track record.

Retrieval

Hybrid RRF fusion (k=60, Cormack 2009) over three signals — full-text, BM25 (ts_rank_cd), and pgvector HNSW — with entropy-routed weighting that shifts from keyword-heavy to semantic as the query's Shannon entropy rises.

Answers arrive in compact by default: abbreviated keys, content truncated at 1200 chars. 30% fewer tokens, and a slightly better nDCG — measured on the 221 id-scored questions, +0.0090 with a paired 95% interval of [+0.0024, +0.0166]. The format genuinely cannot change which documents rank; what it changes is how many of them survive the response token budget before they are scored. Verbose at the default 5000-token budget weighs 5286 tokens and loses its tail; compact weighs 3723 and keeps it. Pass "format": "verbose" for the full per-branch score breakdown.

Verification that actually verifies

cuba_faro mode=verify checks a claim against what is stored. It used to score claims by cosine similarity to the retrieved evidence, and that does not work — similarity measures what a text is about, not what it asserts. "cuba-memorys is written in Rust" and "…in Java" are nearly the same vector. Measured on the live corpus, the false claim scored 0.61 and the true one 0.59.

Entailment is a different question from similarity, and it needs something that reads. A local cross-encoder now judges each piece of evidence — supports / contradicts / unrelated — and confidence is derived from the verdicts, each weighted by that evidence's similarity. Same corpus, after:

Being on-topic is not support, and unrelated counts for neither side.

The judge is mDeBERTa-v3-base-xnli running locally on ONNX: 100 languages, ~50 ms per verdict, no API key, no network, no cost. That matters here — about 75% of this corpus is Spanish, and the English-only NLI models everyone reaches for first would have silently failed on three memories out of four. Install it with cuba-memorys models nli; cuba-memorys doctor will tell you whether it loaded.

Without it, verification falls back to an LLM (your MCP client's own model via sampling, a local claude CLI, or the Anthropic API) — and with none of those, to an honest unknown rather than an invented verdict.

Two things it will not do. It will not confirm a claim on weak evidence: entailment must clear 0.80 while contradiction needs only 0.60, because confirming a false memory and doubting a true one are not errors of equal cost. And when it cannot tell, it says so instead of returning whichever number came out largest — an argmax over a 3-way head will happily publish supports for a claim that is flatly false, and did.

Calibrated abstention

The out-of-distribution gate rejects queries the corpus cannot answer. The threshold is not a magic constant: Ledoit-Wolf covariance shrinkage plus a conformal quantile, calibrated against your own corpus with cuba-memorys calibrate --dataset <questions.jsonl> --apply and persisted (the dataset is required — without it the command refuses). (The theoretical χ² threshold rejected 100% of answerable queries. Distribution-free calibration is not a nicety here.)

Sync between machines, through git

CUBA_MODE=red puts two machines on one database. cuba_sync is the other route, for machines that never see each other: the graph is written out as JSON you can commit, and read back on the other side.

cuba-memorys sync export            # write the bundle under .cuba-memorys/
cuba-memorys sync import            # read one back in
cuba-memorys sync diff              # entities on disk vs entities in the database
cuba-memorys sync status            # which bundles this machine has already imported
cuba-memorys hook install           # export after every commit, import after every checkout

The same four actions are cuba_sync action=export|import|diff|status. A bundle is one JSON file per entity with its observations inside, plus episodes/YYYY-MM/, errors/, decisions/, relations.json, projects.json, tombstones.json and a manifest.json — the active project and anything not bound to a project, unless you pass --scope all. Embeddings stay out unless you ask for them (--with-embeddings): they are most of the bytes and they can be recomputed. A bundle imports once, and the manifest hash covers the contents of every file in it — so an unchanged bundle is skipped, and a hand-edited entity file is a new bundle rather than a silent no-op.

A deletion travels now, and stops where it would take something with it. Deleting a row records a tombstone, and the receiving side deletes exactly the ids that were named. Before this, a delete was not slow to arrive — it was undone: the peer still had the row, exported it, and it came back on the next round trip. The entity tombstone is the dangerous one, because deleting an entity cascades to everything hanging off it. It is applied only when this machine has no observations or episodes under that entity that the sender never named; otherwise it is withheld and reported in tombstones_withheld. A tombstone for an entity with three children there must not take three hundred here.

And a bundle cannot quietly wipe you. If the tombstones in it would delete at least 25 rows and more than 10% of the observations on this machine, the import refuses and asks for confirm=true. A remote wipe and a large legitimate cleanup look identical; the only difference is whether you meant it. The floor matters as much as the ratio: on a database with a single observation a pure percentage demanded confirmation to delete that one, and a guard that trips on ordinary curation is one everybody learns to pass confirm=true through — and then it guards nothing.

conflict=merge does not merge content, and now says so. merge and skip are one policy: rows that are missing here arrive, and where a row already exists with different content, the one that was here first wins and the incoming text is dropped. What changed is the silence — the import counts those rows and reports them as diverged, with their ids and a note saying what it did. conflict=overwrite takes the incoming version and keeps the one it replaced in previous_versions (the newest 20 are kept), and clears the embedding when the content changed, so a row stops being retrievable by a meaning it no longer carries.

Counters do merge, under either policy. importance and access_count on an entity, and strength on a relation, are not values one side copies from the other: each machine grows its own, from its own reinforcement and its own traversals. The higher of the two wins, which is idempotent — importing the same bundle twice inflates nothing. (Summing would be more faithful to "both machines counted", and would double on a re-import, so it loses to a rule that cannot corrupt the number.)

Which machine is which. Each installation generates a uuid in its own database on first migration — one row, stable across restarts, unique by construction — and the manifest carries it, so a bundle can say which machine produced it. CUBA_NODE_NAME keeps meaning what it always meant: a human-readable label stored in origin_node. It is not the identity and could not be one, because two machines both called pop-os is the likeliest outcome there is.

The clock ticks for what a peer needs, and stays still for local noise. An observation's version advances when its content, type, trust, evidence level or tags actually change, and for nothing else. Decay moves importance and last_accessed; reembed replaces vectors. If either woke the clock, every export would ship a graph that had not changed and the two machines would never stop talking to each other about nothing. Rewriting a row with the same content does not tick it either, so an idempotent re-import does not invent a conflict out of agreement.

Older bundles still import. The format is SCHEMA_VERSION 2: version, updated_at, origin_node, previous_versions, evidence, verification and trust travel now, because a conflict rule that compares clocks needs the clock to be in the file. Bundles written before that still import — every new field defaults, and a v1 observation lands as asserted, which is the honest reading of a file that never claimed anything stronger.

Anything in an incoming bundle that looks like a credential is stored quarantined instead of trusted — withheld from cuba_faro and cuba_expediente until you promote it with cuba_eco — because an import reads JSON out of a repository anyone with push access can write to.

And it tells you when it is broken

$ cuba-memorys doctor
[  ok  ] migrations           49 aplicadas, ninguna dirty
[  ok  ] embedding_dim        runtime 1024-d == columna vector(1024)
[  ok  ] runtime_role         'cuba_app' sin superuser — RLS y audit efectivos
[ warn ] binary_freshness     4 proceso(s) MCP corren un binario más viejo que el de disco

This exists because the failure mode of a hybrid search engine is not a crash — it is a vector branch dying and the search quietly becoming lexical, with no symptom. The server now refuses to start on an embedding-dimension mismatch, and search sets degraded: true in the response when a branch fails.

The CLI: your memory without an LLM in the middle

Twenty-two commands. cuba-memorys --help lists them all.

dedupe — because a different string is a different entity

cuba_alma create inserts with ON CONFLICT (name). So one project fragments into Mapupita-Web, Mapupitta-Web (typo), Mapupita Web, mapupita… and searching one finds none of the others. On a real 266-entity graph, 158 of them (59%) had not a single relation — for PageRank and multi-hop retrieval, they did not exist.

What decides a merge is not the embedding centroid. That was the obvious idea and it is wrong: M-Codes Reference Guide and G-Codes Reference Guide sit at 0.811 cosine between centroids. On a corpus about one domain, centroid similarity measures the domain, not the entity — a 0.80 threshold would have merged two different CNC guides, irreversibly.

So --apply merges only what is provable (identical after normalizing case and separators). Typos and near-matches are shown, and judged one at a time with --judge. The old name is written to brain_entity_aliases, so nothing is lost: looking it up still resolves.

The 28 tools

Named after Cuban culture. cuba-memorys advertises all of them, or set CUBA_TOOL_PROFILE=lean to advertise an everyday core of 6 plus cuba_tools + cuba_call — 8 of 30, a 73% smaller catalogue with zero functions lost, the rest reachable on demand.

Knowledge graph — cuba_alma (entities) · cuba_cronica (observations, episodes, timeline) · cuba_puente (typed relations, traversal, link prediction) · cuba_ingesta (bulk import)

Search — cuba_faro (hybrid RRF, verification, MMR diversification, OOD abstention)

Error memory — cuba_alarma (report) · cuba_remedio (resolve) · cuba_expediente (search past errors; warns if an approach failed before)

Sessions & decisions — cuba_jornada (session lifecycle, diff) · cuba_decreto (architecture decisions) · cuba_proyecto (per-project isolation) · cuba_pre_compact (survive /compact)

Procedural — cuba_receta (recipes ranked by Wilson lower bound)

Cognition — cuba_reflexion (gap detection) · cuba_hipotesis (abductive inference) · cuba_contradiccion (semantic conflicts) · cuba_juez (LLM judge) · cuba_centinela (prospective triggers) · cuba_calibrar (Bayesian calibration, source credibility)

Maintenance — cuba_zafra (decay, prune, merge, PageRank, Leiden communities) · cuba_eco (RLHF feedback) · cuba_vigia (health, drift, centrality) · cuba_forget (GDPR erasure) · cuba_archivo (CFR-21 hash-chain audit log) · cuba_pizarra (working memory) · cuba_sync (git-friendly export/import between machines, with propagated deletions and a remote-wipe guard)

Meta — cuba_tools (discover) · cuba_call (invoke)

Configuration

Footprint

A memory server is infrastructure: it is running when you are not using it. On the 6 GB laptop GPU this was measured on, it used to hold 5228 MiB of VRAM from boot — 93% of the card — and other GPU programs stopped being able to start. The NVIDIA driver was returning NV_ERR_NO_MEMORY on channel creation, which is what a game or a GPU-accelerated terminal fails on.

Two of the numbers below ship as defaults; two need a line of config, and this table keeps them apart rather than quoting the best one as if it came free.

Everything in the defaults column is code that ships. CUBA_RERANK_CHUNK=4 is one env var. The last column additionally needs the rebuilt reranker described below. None of it removed a feature.

Four things got it there:

Placement per model, not per process. Only the reranker is accelerated by a GPU — the INT8 embedder cannot be, and the FP32 NLI is not worth a gigabyte of VRAM for a judge that runs occasionally and tolerates 150-400 ms. The arena cap is per session, so three sessions asking for CUDA on a 6 GB card is a 3× overcommit waiting to fail.

A CUDA arena that stops doubling. ArenaExtendStrategy::NextPowerOfTwo is the ONNX Runtime default and it reserves in powers of two rather than what the session asked for.

The reranker loads on its first batch. Under socket activation the daemon starts far more often than it reranks, and plenty of those starts only ever answer a save.

A daemon that is not running when nobody is asking. CUBA_IDLE_SHUTDOWN_SECS plus a systemd .socket unit: the socket owns the port, the daemon starts on the first real connection and exits after the idle window. It shuts down through the normal path — serve returns, the background drain flushes in-flight embedding writes, sqlx closes its pool — because exiting the process directly loses those writes silently.

# ~/.config/systemd/user/cuba-memorys.socket
[Socket]
ListenStream=127.0.0.1:8787
Accept=no

[Install]
WantedBy=default.target

# ~/.config/systemd/user/cuba-memorys.service — no [Install]; the socket starts it
[Unit]
Requires=cuba-memorys.socket

[Service]
Type=exec
ExecStart=%h/.local/bin/cuba-memorys-daemon serve 127.0.0.1:8787
# An idle shutdown exits 0 — Restart=always would bounce it straight back up.
Restart=on-failure
Environment=CUBA_IDLE_SHUTDOWN_SECS=1200
Environment=CUBA_EMBED_DEVICE=cpu
Environment=CUBA_RERANK_DEVICE=gpu
Environment=CUBA_NLI_DEVICE=cpu

Host RAM: it sizes itself to your machine

VRAM was only half of it. The weights also live in host memory, and that appetite used to be fixed no matter what the machine had. Measured with cargo run --release --features cuda --example mem_bench, daemon stopped, on the 6 GB laptop GPU:

Resident settles near 1941 MiB; the peak is 2677 because loading a 1,1 GB ONNX file costs transient memory on top of the weights it leaves behind. The peak is the number that has to fit, not the steady state.

On the machine this was measured on that is fine. On a 4 GB laptop it is not, and under a systemd unit capped at MemoryHigh=4500M it has been seen paging 2,56 GiB to swap — MemoryHigh does not kill, it reclaims, and reclaiming is paging.

Two traps worth knowing if you re-run this. mem_bench attributes VRAM to its own PID via nvidia-smi --query-compute-apps, because reading memory.used charges you for every other process on the card — that is how a first attempt showed 3590 MiB "at process start" that belonged to a game and a desktop shell. And run it with the daemon's own environment: with CUBA_RERANKER_PATH unset it silently loads the unfused artifact and the warm-up goes from 3,7 s to 131 s on CPU.

So the daemon now reads the machine at startup and picks a level. Nothing is invented for this: all three degradations already existed and are tested.

How the level is chosen. The budget is min(cgroup limit, system available) − 768 MiB of headroom, and the cgroup has to win. On this machine /proc/meminfo reports 7,16 GB available while the daemon's cgroup caps it at 4,39 GiB — believing /proc would load 2,6 GiB of weights against a limit where the kernel already starts paging. The reader walks from the cgroup root down to the leaf and takes the tightest memory.max or memory.high it finds, because the limit is usually set on an ancestor.

Models are then fitted in order of measured value: the embedder first, then the reranker (+93% nDCG, so it outranks the judge), then NLI.

The plan can only take away. Every knob is capped at the value the daemon already used, so on a machine with room the level is full and nothing changes. Degradation only goes downward.

You always win. Any of these set by hand is left untouched — the regulator fills gaps, it does not overwrite decisions:

CUBA_EMBED_INTRA_THREADS   CUBA_RERANK_INTRA_THREADS   CUBA_NLI_INTRA_THREADS
CUBA_RERANK_CHUNK          CUBA_GPU_MEM_LIMIT_MB       CUBA_OOD_FIT_LIMIT
CUBA_DB_MAX_CONNECTIONS

To force a model off regardless of the budget, point it at a path that does not exist — CUBA_RERANKER_PATH=/nonexistent or CUBA_NLI_PATH=/nonexistent. That is the same mechanism the regulator itself uses.

To see what it decided, run cuba-memorys doctor: it reports the reading and the resulting plan, and warns when the level falls to minimal. The plan is also logged at startup with the full machine reading behind it.

The reranker artifact

The published bge-reranker-v2-m3 ONNX is converted to FP16 before any graph fusion, which leaves 785 Cast nodes threaded through it. ONNX Runtime claws some of that back at load time (2023 → 897 nodes, 49 SkipLayerNormalization), but it cannot fuse Gelu and it repeats the work on every cold start. Rebuilding from the FP32 export and fusing first:

python -m onnxruntime.transformers.optimizer \
  --input model.onnx --output model.onnx \
  --model_type bert --num_heads 16 --hidden_size 1024 \
  --opt_level 1 --use_gpu --float16

Identical top-10 order on a real search, fused_score differing by at most 0,0029; on synthetic logits at the real batch shapes, Pearson ≥ 0,9997 with the same ranking in every batch.

Attention does not fuse, and that is not fixable here. is_fully_optimized: Attention (or MultiHeadAttention) not fused, at opt_level 0, 1, 2 and 99, on both the FP16 artifact and the clean FP32 one. The export builds its Q/K/V reshapes from dynamic shape subgraphs (Shape → Gather → Unsqueeze → Concat → Reshape) and AttentionFusion needs a Reshape with a constant shape to read num_heads and head_size off it. So flash/efficient attention stays unavailable without a re-export using static shapes — worth knowing before anyone spends an afternoon on it.

Measured — and the benchmark that was lying

Until v0.12 this section carried a line reading "every number here is measured rather than assumed", and every number in it was wrong. The benchmark was broken in three ways, and finding out cost two published conclusions.

It had ten queries. A 95% interval of roughly ±0.12; the smallest effect it could detect was ~0.25 nDCG. Any claim about a smaller difference was noise wearing a decimal point.

Relevance was judged by substring match. A result counted as correct if its text merely contained a marker word — so every observation mentioning "postgres" scored as a right answer to any question about postgres, whether it answered anything or not. That measures keyword presence, not retrieval, and it tilts the whole benchmark toward the lexical branch and against the vector one.

nDCG normalized against what was retrieved, not what exists. With 5 relevant documents in the corpus and 2 found, the "ideal" ranking was taken to be those 2 — so a system that missed 60% of the answer scored a perfect 1.0. (And R@10 = 3.125 shipped in this file. Recall is a proportion.)

The real number is not 0.894. On 221 id-scored queries it is nDCG@10 = 0.50 [95% CI 0.44–0.56]. The system did not get worse. It was never 0.894.

What that cost

• "The cross-encoder reranker earns nothing" — it had never run. Three bugs in series: faro wrapped the call in if let Ok(..) and dropped the error; it fed token_type_ids to a model that is XLM-RoBERTa and has none; it read f16 logits as f32. The output was "bit for bit identical" to no reranking not because reranking changed nothing, but because it never happened. Fixed; being measured properly now.

• Associative retrieval does degrade — but the old evidence (−0.03 at n=10) could not have shown it. On the new dataset with a paired bootstrap (the correct test: same queries in both arms), the interval is [−0.051, −0.018] and never touches zero. It improves 0 queries and hurts 23. The decision was right; the reasoning was not. The power was never in more data — it was in using the right test.

What survives, re-measured honestly

Foundations

Development

git clone https://github.com/LeandroPG19/cuba-memorys.git
cd cuba-memorys/rust && cargo build --release

# On an NVIDIA machine, build this way instead — without it the reranker spends
# its whole budget for a ranking that gets discarded. It accelerates the
# reranker only; see Footprint for why the other two models stay on the CPU.
cargo build --release --features docs,cuda

./scripts/demo.sh                  # runs on a throwaway Postgres it removes on exit
./scripts/merge-gate.sh            # fmt · clippy -D warnings · 316 tests · audit · integration
cargo run --release --example rerank_bench   # does the reranker fit its budget here?

Publishing is tag-driven: v* triggers GitHub Release binaries (5 platforms), PyPI wheels, npm, and the MCP Registry. A test pins all four files that hold a version number to the same value, because they used to drift and nothing caught it.

License

Apache-2.0 — use it, modify it, ship it, sell it, embed it in a closed product. No copyleft obligation. The licence also grants patent rights explicitly, which is the part legal departments care about.

Author

Leandro Perez G. — @LeandroPG19