Cuba-Memorys

PyPI npm MCP Registry Rust PostgreSQL License: CC BY-NC 4.0 Audit Tech Debt

Persistent memory for AI agents — A Model Context Protocol (MCP) server that gives AI coding assistants long-term memory with a knowledge graph, neuroscience-inspired algorithms, and anti-hallucination grounding.

19 tools with Cuban soul. Sub-millisecond handlers. Mathematically rigorous.

IMPORTANT

v0.7.0 — 10 algorithmic improvements + 19 bug fixes + comprehensive audit. PageRank blend (preserves Hebbian/BCM learned importance), hybrid verify (trigram + embedding fusion), ONNX concurrency semaphore (Little's Law), sigmoid entropy routing (replaces step function), word-level session boost, weighted Hebbian neighbor diffusion, exponential coverage saturation, O(n) entropy counting. Fixed: hash embeddings corrupting DB, centrality normalization, cache LRU, jornada race condition, 6 MCP schemas. Removed blake3 dependency. 68 tests, 0 clippy warnings, 0 tech debt.

Demo

Why Cuba-Memorys?

AI agents forget everything between conversations. Cuba-Memorys solves this:

Stratified exponential decay — Memories fade by type (facts=30d, errors=14d, context=7d), strengthen with access
Hebbian + BCM metaplasticity — Self-normalizing importance via Oja's rule with EMA sliding threshold
Hybrid RRF fusion search — pg_trgm + full-text + pgvector HNSW, entropy-routed weighting (k=60), temporal filters, tag filters, compact format
Knowledge graph — Entities, observations, typed relations with Leiden community detection and Adamic-Adar link prediction
Anti-hallucination grounding — Verify claims with graduated confidence + Bayesian calibration over time
Episodic memory — Separate temporal events (Tulving 1972) with power-law decay I(t) = I₀/(1+ct)^β (Wixted 2004)
Contradiction detection — Scan for semantic conflicts via embedding cosine + bilingual negation heuristics
Prospective memory — Triggers that fire on entity access, session start, or error match ("remind me when X")
Contextual Retrieval — Entity context prepended before embedding (Anthropic technique, +20% recall)
REM Sleep consolidation — Autonomous stratified decay + PageRank + auto-prune + auto-merge + episode decay
Graph intelligence — PageRank, Leiden communities, Brandes centrality, Shannon entropy, gap detection
Session awareness — Provenance tracking, session diff, importance priors per observation type
Error memory — Never repeat the same mistake (anti-repetition guard + pattern detection)

Comparison

Feature	Cuba-Memorys	Basic Memory MCPs
Knowledge graph with typed relations	Yes	No
Exponential importance decay	Yes	No
Hebbian learning + BCM metaplasticity	Yes	No
Hybrid entropy-routed RRF fusion	Yes	No
KG-neighbor query expansion	Yes	No
GraphRAG topological enrichment	Yes	No
Leiden community detection	Yes	No
Brandes betweenness centrality	Yes	No
Shannon entropy analytics	Yes	No
Adaptive prediction error gating	Yes	No
Anti-hallucination verification	Yes	No
Error pattern detection	Yes	No
Session-aware search boost	Yes	No
REM Sleep autonomous consolidation	Yes	No
Multilingual ONNX embeddings (e5-small)	Yes	No
Episodic memory (power-law decay)	Yes	No
Contradiction detection	Yes	No
Prospective memory triggers	Yes	No
Bayesian confidence calibration	Yes	No
Link prediction (Adamic-Adar)	Yes	No
Auto-tagging (TF-IDF)	Yes	No
Contextual Retrieval (Anthropic)	Yes	No
Temporal search filters	Yes	No
Zero-config Docker auto-setup	Yes	No
Write-time dedup gate	Yes	No
Contradiction auto-supersede	Yes	No
GDPR Right to Erasure	Yes	No
Graceful shutdown (SIGTERM/SIGINT)	Yes	No

Installation

PyPI (recommended)

pip install cuba-memorys

npm

npm install -g cuba-memorys

From source

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

Binary download

Pre-built binaries available at GitHub Releases.

Quick Start

Zero configuration required — just install and add to your editor. Cuba-memorys automatically provisions a PostgreSQL database via Docker on first run.

Prerequisite: Docker must be installed and running.

npm install -g cuba-memorys
claude mcp add cuba-memorys -- cuba-memorys

That's it. On first run, Cuba-memorys will:

Detect that no database is configured
Create a Docker container with PostgreSQL + pgvector
Initialize the schema automatically
Start serving 19 MCP tools

npm install -g cuba-memorys

Add to your MCP config (.cursor/mcp.json, .windsurf/mcp.json, or .vscode/mcp.json):

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

No DATABASE_URL needed — auto-provisioned via Docker on first run.

If you already have PostgreSQL with pgvector, set the environment variable:

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

Optional: Multilingual ONNX Embeddings

For real multilingual-e5-small semantic embeddings (94 languages, 384d) instead of hash-based fallback:

./rust/scripts/download_model.sh  # Downloads ~113MB model
export ONNX_MODEL_PATH="$HOME/.cache/cuba-memorys/models"
export ORT_DYLIB_PATH="/path/to/libonnxruntime.so"

Without ONNX, the server uses deterministic hash-based embeddings — functional but without semantic understanding. With ONNX, Contextual Retrieval prepends [entity_type:entity_name] to content before embedding for +20% recall.

The 19 Tools

Every tool is named after Cuban culture — memorable, professional, meaningful.

Knowledge Graph

Tool	Meaning	What it does
`cuba_alma`	Alma — soul	CRUD entities. Types: `concept`, `project`, `technology`, `person`, `pattern`, `config`. Hebbian boost + access tracking. Fires prospective triggers on access.
`cuba_cronica`	Cronica — chronicle	Observations with semantic dedup, PE gating V5.2, importance priors by type, auto-tagging (TF-IDF top-5 keywords), session provenance, contextual embedding. Also manages episodic memories (episode_add/episode_list) and timeline view.
`cuba_puente`	Puente — bridge	Typed relations. Traverse walks the graph. Infer discovers transitive paths. Predict suggests missing relations via Adamic-Adar link prediction.
`cuba_ingesta`	Ingesta — intake	Bulk knowledge ingestion: accepts arrays of observations or long text with auto-classification by paragraph.

Search & Verification

Tool	Meaning	What it does
`cuba_faro`	Faro — lighthouse	RRF fusion (k=60) with sigmoid entropy routing, pgvector, temporal filters (`before`/`after`), tag filters, score breakdown (text/vector/importance/session), compact format (~35% fewer tokens), hybrid verify (trigram + embedding fusion), Bayesian calibrated accuracy, token-budget truncation, `max_tokens` control.

Error Memory

Tool	Meaning	What it does
`cuba_alarma`	Alarma — alarm	Report errors. Auto-detects patterns (>=3 similar = warning). Fires prospective triggers on error match.
`cuba_remedio`	Remedio — remedy	Resolve errors with cross-reference to similar unresolved issues.
`cuba_expediente`	Expediente — case file	Search past errors. Anti-repetition guard: warns if similar approach failed before.

Sessions & Decisions

Tool	Meaning	What it does
`cuba_jornada`	Jornada — workday	Session tracking with goals, outcomes, session diff (what was learned), and previous session context on start. Fires prospective triggers.
`cuba_decreto`	Decreto — decree	Record architecture decisions with context, alternatives, rationale.

Cognition & Analysis

Tool	Meaning	What it does
`cuba_reflexion`	Reflexion — reflection	Gap detection: isolated entities, underconnected hubs, type silos, observation gaps, density anomalies (z-score).
`cuba_hipotesis`	Hipotesis — hypothesis	Abductive inference: given an effect, find plausible causes via backward causal traversal. Plausibility = path_strength x importance.
`cuba_contradiccion`	Contradiccion — contradiction	Scan for semantic conflicts between same-entity observations via embedding cosine + bilingual negation heuristics.
`cuba_centinela`	Centinela — sentinel	Prospective memory triggers: "remind me when X is accessed / session starts / error matches". Auto-deactivate on max_fires, expiration support.
`cuba_calibrar`	Calibrar — calibrate	Bayesian confidence calibration: track faro/verify predictions, compute P(correct\|grounding_level) via Beta distribution. Closes the verify-correct feedback loop.

Memory Maintenance

Tool	Meaning	What it does
`cuba_zafra`	Zafra — sugar harvest	Stratified decay (30d/14d/7d by type), power-law episode decay, prune, merge, summarize, pagerank, find_duplicates, export, stats, reembed (model migration with versioning). Auto-consolidation on >50 observations.
`cuba_eco`	Eco — echo	RLHF feedback: positive (Oja boost), negative (decrease), correct (update with versioning).
`cuba_vigia`	Vigia — watchman	Analytics: summary, enhanced health (null embeddings, active triggers, table sizes, embedding model), drift (chi-squared), Leiden communities, Brandes bridges.
`cuba_forget`	Forget — forget	GDPR Right to Erasure: cascading hard-delete of entity and ALL references (observations, episodes, relations, errors, sessions). Irreversible.

Architecture

cuba-memorys/
├── docker-compose.yml           # Dedicated PostgreSQL 18 (port 5488)
├── server.json                  # MCP Registry manifest
├── pyproject.toml               # Maturin (bindings = "bin") — PyPI wheel
├── package.json                 # npm wrapper
└── rust/                        # v0.7.0
    ├── src/
    │   ├── main.rs              # mimalloc + graceful shutdown (SIGTERM/SIGINT)
    │   ├── lib.rs               # Shared types and utilities
    │   ├── protocol.rs          # JSON-RPC 2.0 + REM daemon (4h cycle)
    │   ├── db.rs                # sqlx PgPool (10 max, 600s idle, 1800s lifetime)
    │   ├── setup.rs             # Zero-config Docker PostgreSQL auto-provisioning
    │   ├── schema.sql           # 8 tables, 20+ indexes, HNSW
    │   ├── constants.rs         # Tool definitions, thresholds, importance priors
    │   ├── handlers/            # 19 MCP tool handlers (1 file each)
    │   ├── cognitive/           # Hebbian/BCM (hebbian.rs), PE gating V5.2
    │   │                        # (prediction_error.rs), dual-strength learning
    │   │                        # (dual_strength.rs), Shannon entropy (density.rs)
    │   ├── search/              # RRF fusion (rrf.rs), confidence (confidence.rs),
    │   │                        # LRU cache (cache.rs)
    │   ├── graph/               # Brandes centrality (centrality.rs), Leiden
    │   │                        # communities (community.rs), PageRank (pagerank.rs)
    │   └── embeddings/          # ONNX multilingual-e5-small (contextual embedding,
    │                            # spawn_blocking, hash fallback)
    ├── scripts/
    │   └── download_model.sh    # Download multilingual-e5-small ONNX (~113MB)
    └── tests/                   # 55 unit + 13 smoke tests

Performance: Rust vs Python

Metric	Python v1.6.0	Rust v0.7.0
Binary size	~50MB (venv)	7.6MB
Entity create	~2ms	498us
Hybrid search	<5ms	2.52ms
Analytics	<2.5ms	958us
Memory usage	~120MB	~15MB
Startup time	~2s	<100ms
Dependencies	12 Python packages	0 runtime deps

Database Schema

Table	Purpose	Key Features
`brain_entities`	KG nodes	tsvector + pg_trgm + GIN indexes, importance, bcm_theta
`brain_observations`	Facts with provenance	9 types, versioning, `vector(384)`, importance priors, auto-tags TEXT[], session_id FK, embedding_model tracking
`brain_relations`	Typed edges	5 types, bidirectional, Hebbian strength, ON CONFLICT dedup
`brain_errors`	Error memory	JSONB context, synapse weight, pattern detection
`brain_sessions`	Working sessions	Goals (JSONB), outcome tracking, session diff
`brain_episodes`	Episodic memory	Tulving 1972, actors/artifacts TEXT[], power-law decay (Wixted 2004)
`brain_triggers`	Prospective memory	on_access/on_session_start/on_error_match, max_fires, expiration
`brain_verify_log`	Bayesian calibration	claim, confidence, grounding_level, outcome (correct/incorrect)

Search Pipeline

Reciprocal Rank Fusion (RRF, k=60) with entropy-routed weighting:

#	Signal	Source	Condition
1	Entities (ts_rank + trigrams + importance)	`brain_entities`	Always
2	Observations (ts_rank + trigrams + importance)	`brain_observations`	Always
3	Errors (ts_rank + trigrams + synapse_weight)	`brain_errors`	Always
4	Vector cosine distance (HNSW)	`brain_observations.embedding`	pgvector installed
5	Episodes (ts_rank + trigrams + importance)	`brain_episodes`	Always

Post-fusion pipeline: Dedup -> KG-neighbor expansion -> Session boost -> Score breakdown -> GraphRAG enrichment -> Token-budget truncation -> Compact format (optional) -> Batch access tracking

Filters: before/after (ISO8601 temporal), tags (keyword), format (verbose/compact)

Mathematical Foundations

Built on peer-reviewed algorithms, not ad-hoc heuristics:

Stratified Exponential Decay (V4)

importance_new = importance * exp(-0.693 * days_since_access / halflife)

Stratified by observation type: facts/preferences=30d, errors/solutions=14d, context/tool_usage=7d. Decision/lesson observations are protected (never decay). Episodic memories use power-law: I(t) = 0.5 / (1 + 0.1*t)^0.5 (Wixted 2004). Importance directly affects search ranking (score0.7 + importance0.3).

Hebbian + BCM — Oja (1982), Bienenstock-Cooper-Munro (1982)

Positive: importance += eta * throttle(access_count, theta_M)
BCM EMA: theta_M = max(10, (1-alpha)*theta_prev + alpha*access_count)

V3: theta_M persisted in bcm_theta column for true temporal smoothing.

RRF Fusion — Cormack (2009)

RRF(d) = sum( w_i / (k + rank_i(d)) )   where k = 60

Entropy-routed weighting via smooth sigmoid (Jaynes 1957 MaxEnt):

t = sigmoid(2.0 * (entropy - 2.75))
text_w  = 0.7 - 0.4*t    (keyword-heavy → balanced → semantic-heavy)
vector_w = 0.3 + 0.4*t   (always sums to 1.0)

Replaces V2 step function which had 40% relative jumps at thresholds.

PageRank Blend — Brin & Page (1998)

importance_new = 0.3 * rank_normalized + 0.7 * importance_existing

Min-max normalized ranks blended via convex combination (α=0.3). Preserves Hebbian/BCM/RLHF accumulated importance instead of overwriting. Uniform distribution guard: skips blend when all ranks are equal (no structural signal).

Other Algorithms

Algorithm	Reference	Used in
Leiden communities	Traag et al. (Nature 2019)	`community.rs` -> `vigia.rs`
PageRank + blend	Brin & Page (1998)	`pagerank.rs` -> convex combination (α=0.3)
Brandes centrality	Brandes (2001)	`centrality.rs` -> undirected normalization
Adaptive PE gating	Friston (Nature 2023)	`prediction_error.rs` -> `cronica.rs`
Shannon entropy	Shannon (1948)	`density.rs` -> information gating
Chi-squared drift	Pearson (1900)	Error distribution change detection
Power-law forgetting	Wixted (2004)	`setup.rs` -> episodic memory decay
Contextual Retrieval	Anthropic (2024)	`onnx.rs` -> entity context prepend
Adamic-Adar	Adamic & Adar (2003)	`puente.rs` -> link prediction
Episodic/Semantic	Tulving (1972)	`brain_episodes` vs `brain_observations`
Bayesian calibration	Beta distribution	`calibrar.rs` -> P(correct\|level)

Configuration

Environment Variables

Variable	Default	Description
`DATABASE_URL`	—	PostgreSQL connection string (auto-provisioned via Docker if not set)
`ONNX_MODEL_PATH`	—	Path to multilingual-e5-small model directory (optional)
`ORT_DYLIB_PATH`	—	Path to libonnxruntime.so (optional)
`RUST_LOG`	`cuba_memorys=info`	Log level filter

Docker Compose

Dedicated PostgreSQL 18 Alpine:

Port: 5488 (avoids conflicts with 5432/5433)
Resources: 256MB RAM, 0.5 CPU
Restart: always
Healthcheck: pg_isready every 10s

How It Works

1. The agent learns from your project

Agent: FastAPI requires async def with response_model.
-> cuba_alma(create, "FastAPI", technology)
-> cuba_cronica(add, "FastAPI", "All endpoints must be async def with response_model")

2. Error memory prevents repeated mistakes

Agent: IntegrityError: duplicate key on numero_parte.
-> cuba_alarma("IntegrityError", "duplicate key on numero_parte")
-> cuba_expediente: Similar error found! Solution: "Add SELECT EXISTS before INSERT"

3. Anti-hallucination grounding

Agent: Let me verify before responding...
-> cuba_faro("FastAPI uses Django ORM", mode="verify")
-> confidence: 0.0, level: "unknown" — "No evidence. High hallucination risk."

4. Memories decay naturally

Initial importance:    0.5  (new observation)
After 30d no access:  0.25 (halved by exponential decay)
After 60d no access:  0.125
Active access resets the clock — frequently used memories stay strong.

5. Community intelligence

-> cuba_vigia(metric="communities")
-> Community 0 (4 members): [FastAPI, Pydantic, SQLAlchemy, PostgreSQL]
  Summary: "Backend stack: async endpoints, V2 validation, 2.0 ORM..."
-> Community 1 (3 members): [React, Next.js, TypeScript]
  Summary: "Frontend stack: React 19, App Router, strict types..."

Security & Audit

Internal Audit Verdict: GO (2026-03-28)

Check	Result
SQL injection	All queries parameterized (sqlx bind)
SEC-002 wildcard injection	Fixed (POSITION-based)
CVEs in dependencies	0 active (sqlx 0.8.6, tokio 1.50.0)
UTF-8 safety	`safe_truncate` on all string slicing
Secrets	All via environment variables
Division by zero	Protected with `.max(1e-9)`
Error handling	All `?` propagated with `anyhow::Context`
Clippy	0 warnings
Tests	68 passing (55 unit + 13 smoke) + 49 E2E
Licenses	All MIT/Apache-2.0 (0 GPL/AGPL)

Dependencies

Crate	Purpose	License
`tokio`	Async runtime	MIT
`sqlx`	PostgreSQL (async)	MIT/Apache-2.0
`serde` / `serde_json`	Serialization	MIT/Apache-2.0
`pgvector`	Vector similarity	MIT
`ort`	ONNX Runtime (optional)	MIT/Apache-2.0
`tokenizers`	HuggingFace tokenizers	Apache-2.0
`mimalloc`	Global allocator	MIT
`tracing`	Structured JSON logging	MIT
`lru`	O(1) LRU cache	MIT
`chrono`	Timezone-aware timestamps	MIT/Apache-2.0

Version History

Version	Key Changes
0.7.0	10 algorithmic improvements: PageRank blend (α=0.3, preserves Hebbian/BCM), hybrid verify (trigram+embedding fusion), ONNX semaphore (Little's Law), sigmoid entropy routing (Jaynes MaxEnt), word-level session boost, weighted Hebbian neighbors (Collins & Loftus), exponential coverage saturation, O(n) entropy. 19 bug fixes: hash embeddings corrupting DB (×5), centrality /2, cache LRU, jornada TOCTOU, alarma self-match, 6 schema mismatches. Removed blake3 dependency. MCP Registry publish fixed. npm postinstall version sync. 68 tests, 0 clippy, 0 tech debt.
0.6.0	Contextual Retrieval (+20% recall), importance priors, score breakdown, compact format (~35% fewer tokens), session provenance/diff, semantic dedup, auto-tagging (TF-IDF), Adamic-Adar link prediction, bulk ingest (cuba_ingesta), enhanced health metrics, partial indexes, embedding model versioning. Auto Docker PostgreSQL setup. 19 tools, 56 tests.
0.5.0	Temporal reasoning (before/after/timeline), contradiction detection (cosine + negation heuristics), prospective memory triggers (centinela), Bayesian calibration (calibrar), abductive inference (hipotesis), gap detection (reflexion). 18 tools.
0.4.0	Multilingual embeddings (e5-small, 94 languages), episodic memory (Tulving 1972, power-law Wixted 2004), stratified decay (30d/14d/7d by type), E2E tests in CI with PostgreSQL. 15 tools.
0.3.0	Deep Research V3: exponential decay replaces FSRS-6, dead code eliminated, SEC-002 fix, embeddings storage on write, GraphRAG CTE fix. 13 tools.
0.2.0	Complete Rust rewrite. BCM metaplasticity, Leiden communities, Shannon entropy, blake3 dedup.
1.0-1.6	Python era: 12 tools, Hebbian learning, GraphRAG, REM Sleep, token-budget truncation.

License

CC BY-NC 4.0 — Free to use and modify, not for commercial use.

Author

Leandro Perez G.

GitHub: @LeandroPG19
Email: leandropatodo@gmail.com

Credits

Mathematical foundations: Oja (1982), Bienenstock, Cooper & Munro (1982, BCM), Cormack (2009, RRF), Brin & Page (1998, PageRank), Traag et al. (2019, Leiden), Brandes (2001), Shannon (1948), Pearson (1900, chi-squared), Friston (2023, PE gating), Tulving (1972, episodic memory), Wixted (2004, power-law forgetting), Adamic & Adar (2003, link prediction), Anthropic (2024, Contextual Retrieval), Wang et al. (2022, E5 embeddings), Malkov & Yashunin (2018, HNSW), Jaynes (1957, MaxEnt sigmoid routing), Robertson (1977, score fusion), Collins & Loftus (1975, spreading activation).