pgmnemo

Multi-agent memory substrate for PostgreSQL — provenance-gated, vector-hybrid recall.

License Version PGXN CI PostgreSQL LoCoMo recall@10 LongMemEval recall@10

v0.4.0 (2026-05-15): hybrid retrieval promoted to default. LoCoMo session recall@10 +4.15pp (0.7951 → 0.8409, p_corr=0.0156), MRR +7.96pp (p_corr<0.0001). 5 significant improvements, 0 regressions across 24 LoCoMo cells. LongMemEval remains neutral — hybrid does NOT close the BM25 gap (BM25 = 0.982; pgmnemo dense = 0.933). See release scorecard and honest competitive reality.

Opt-out: SET pgmnemo.disable_hybrid = 'true' restores strict v0.3.0 vector-only behaviour.

What’s next: v0.4.1 (target 2026-06-01) — deprecate dead graph machinery from default recall_lessons() path. Full plan: ROADMAP.md.

Benchmarks (v0.3.0, retrieval-only)

Read this before the numbers below: docs/COMPETITIVE_REALITY.md explains exactly what these recall@K figures mean, what they don’t, and where our methodology has asymmetries vs paper baselines and competitor positioning.

Real numbers vs published academic benchmarks. Canonical protocol: docs/BENCHMARK_PROTOCOL.md (v1, frozen 2026-05-13). Full per-version history: benchmarks/METRICS_BY_VERSION.md. Reproduction commands in docs/BENCHMARKS.md.

Benchmark Methodology Embedder recall@10 / MRR Honest comparison
LoCoMo (Maharana ACL 2024) session-level (paper-canonical headline) DRAGON 0.7994 / 0.5569 Easier task than paper Table 3 (272 sessions vs 5882 turns search space)
LoCoMo same paper, turn-level (apples-to-apples with paper baseline) turn-level (retrieval primitive) DRAGON recall@5 = 0.302 / MRR = 0.237 Paper DRAGON dense recall@5 ≈ 0.225 → pgmnemo +7.7pp
LongMemEval-S (Wu ICLR 2025) retrieval-only, full session bge-m3 (subst. for Stella V5)¹ 0.9334 / 0.8472 Loses to BM25 baseline² 0.982 by ~5pp

¹ Stella V5 paper-canonical incompatible with transformers 5.8 — substituted bge-m3 (1024d, MTEB-strong). ² Pure-Python BM25 baseline included for reference: run_nollm.py. On the keyword-heavy LongMemEval workload it currently outperforms our dense vector path. Fixing this is the explicit gate for v0.4.0 — see ROADMAP.md.

The “we beat everyone” framing is wrong. Our headline session-level LoCoMo number compares to a 22× smaller search space than the paper baseline. Our LongMemEval number is below a 50-LOC BM25 script. Comparisons with Mem0 / Zep / MAGMA on these datasets are apples-to-oranges — they optimise different objectives. The honest competitive position is detailed in COMPETITIVE_REALITY.md §3-§5.

Reproduce in 3 commands: see docs/BENCHMARKS.md#reproducibility.

What pgmnemo’s bench does NOT measure (§2 of COMPETITIVE_REALITY.md): insertion throughput, concurrent read/write, retrieval latency p50/p95/p99, multi-tenant RLS correctness, scale beyond ~5k rows, end-to-end agent task completion, provenance gate correctness, state-machine transitions.

Why this exists

  • One differentiator none of Pinecone, Letta, Mem0, or Zep have: a write-time provenance gate. Every ingest() call must carry a commit_sha or artifact_hash; rows without provenance are blocked (or warned) by default. Hallucinated agent memories cannot silently accumulate.
  • No new service. CREATE EXTENSION pgmnemo; in your existing PostgreSQL — no separate API server, no SaaS endpoint, no vendor lock-in.
  • Hybrid recall in-database. Cosine similarity (HNSW) + BM25 full-text + recency decay + importance weighting, scored in one SQL call.
  • Role isolation built in. First-class role + project_id composite scoping; no hand-rolled RLS.
Aspect pgmnemo Generic Vector DB Cloud Memory API
Provenance enforcement ✅ Mandatory ❌ ❌
Zero data egress ✅ In-database ❌ ❌
Install model CREATE EXTENSION External service SaaS API
Self-hosted price Free (Apache 2.0) $$$$ $$$$$

30-second quickstart

PGXN install (if pgxnclient is available):

pgxn install pgmnemo

From source (Docker):

# 1. Start PG 17 + pgvector
docker run --name pgmnemo-dev -e POSTGRES_PASSWORD=pass -p 5432:5432 -d pgvector/pgvector:pg17

# 2. Build and install the extension (requires make, gcc, pg_config on PATH)
git clone https://github.com/pgmnemo/pgmnemo.git
docker exec pgmnemo-dev bash -c "apt-get install -y postgresql-server-dev-17 make gcc 2>/dev/null; true"
docker cp pgmnemo/extension pgmnemo-dev:/tmp/pgmnemo
docker exec pgmnemo-dev bash -c "cd /tmp/pgmnemo && make && make install"

-- psql -h localhost -U postgres

CREATE EXTENSION pgmnemo CASCADE;

SELECT pgmnemo.ingest(
    p_role        := 'developer',
    p_project_id  := 1,
    p_topic       := 'auth',
    p_lesson_text := 'Rotate JWT secrets after any key-compromise incident.',
    p_commit_sha  := 'abc1234'
);

SELECT lesson_text, score
FROM pgmnemo.recall_lessons(
    query_embedding := array_fill(0, ARRAY[1024])::vector(1024),
    query_text      := 'JWT secret rotation',
    role_filter     := 'developer'
);

For a native install (no Docker), see INSTALL.md.

Features

  • HNSW vector search — fast approximate nearest-neighbour recall via pgvector HNSW indexes
  • Provenance gate — enforce / warn / off modes; controlled by pgmnemo.gate_strict GUC
  • Recency-weighted scoring — 0.5×cosine + 0.2×importance + γ×recency(90d) + 0.1×prov_strength; γ tunable via pgmnemo.recency_weight
  • Role scoping — role + project_id composite isolation; role_filter=NULL pools across roles
  • Graph traversal — traverse_causal_chain() and traverse_temporal_window() walk typed mem_edge relationships between lessons
  • MAGMA edge taxonomy (v0.3.0, EXPERIMENTAL) — edge_kind ENUM (semantic | temporal | causal | entity) with per-kind partial indexes; recall_lessons() BFS graph-proximity now correctly uses edge_kind instead of the broken v0.2.x relation_type string matching. MAGMA §4 (adaptive traversal policy) and §5 (dual-stream consolidation) are not yet implemented.

Compatibility

PostgreSQL Status pgvector Platform
17 Fully tested ≥ 0.7.0 required amd64 (Docker + native)
14–16 Best-effort ≥ 0.7.0 required amd64 (Docker + native)
< 14 Not supported — —
arm64 Source-build only ≥ 0.7.0 required No pre-built images

Documentation

License

Apache License 2.0 — see LICENSE.

Contributing

See CONTRIBUTING.md. Contributions accepted under the DCO sign-off model.

Citing

@misc{gaydabura2026pgmnemo,
  author = {Gaydabura, Alex and pgmnemo contributors},
  title  = {pgmnemo: A Provenance-Gated Multi-Agent Memory Substrate for PostgreSQL},
  year   = {2026},
  note   = {ICSE-SEIP submission in preparation}
}