🌿 pgmnemo 0.7.2 — memory that lives *in* Postgres
Most agent memory means standing up a separate service (mem0 / Zep / Letta) next to your database. pgmnemo is a Postgres extension: memory sits right beside pgvector, queryable in plain SQL, learning which lessons actually worked — and letting the rest decay.
0.7.2 makes that one line:
It's a packaging fix — and we added a CI gate that installs the built artifact into a clean Postgres and runs
No new service. No new query language. Just Postgres that remembers.
PGXN · PyPI · GitHub → pgmnemo.com
Most agent memory means standing up a separate service (mem0 / Zep / Letta) next to your database. pgmnemo is a Postgres extension: memory sits right beside pgvector, queryable in plain SQL, learning which lessons actually worked — and letting the rest decay.
0.7.2 makes that one line:
CREATE EXTENSION pgmnemo CASCADE;It's a packaging fix — and we added a CI gate that installs the built artifact into a clean Postgres and runs
CREATE EXTENSION before every release. So "it installs" is now proven, not assumed.No new service. No new query language. Just Postgres that remembers.
PGXN · PyPI · GitHub → pgmnemo.com
🌿 pgmnemo 0.7.2 is live
Agent memory that lives inside Postgres — no separate service to run, no new query language. Your memory sits next to pgvector, queryable in plain SQL, and learns which lessons actually worked: verified ones strengthen, the rest decay.
Install — one line:
In 0.7.2: a packaging fix so it installs cleanly from PGXN, Docker, and source in one step. We also added a release gate that installs the built artifact into a clean Postgres and runs
On 0.7.1?
Postgres that remembers. → pgmnemo.com
GitHub · PGXN · PyPI
Agent memory that lives inside Postgres — no separate service to run, no new query language. Your memory sits next to pgvector, queryable in plain SQL, and learns which lessons actually worked: verified ones strengthen, the rest decay.
Install — one line:
CREATE EXTENSION pgmnemo CASCADE;In 0.7.2: a packaging fix so it installs cleanly from PGXN, Docker, and source in one step. We also added a release gate that installs the built artifact into a clean Postgres and runs
CREATE EXTENSION *before every publish* — so "it installs" is proven, not assumed.On 0.7.1?
ALTER EXTENSION pgmnemo UPDATE — no schema changes.Postgres that remembers. → pgmnemo.com
GitHub · PGXN · PyPI
A known gap in RAG and agent memory: retrieval ranks by similarity, but similarity isn't usefulness. A vector store keeps returning a note that looks relevant and has been wrong every time you used it — because nothing connects "what got retrieved" to "did it actually help." No feedback loop.
Teams patch this by hand: curate memories, prune stale ones, hard-code priorities. Doesn't scale, isn't auditable.
pgmnemo (a Postgres extension for agent memory) puts the feedback loop in the database. Each stored memory — it calls them "lessons" — has a confidence score the retriever reads. After a task, you record the outcome in one line:
SELECT pgmnemo.reinforce(lesson_id, 'success'); -- or 'failure'
Confidence rises for what helped, falls faster for what kept preceding failures. Good track records surface first; dead weight sinks. No extra service, no LLM on the write path.
And because it's Postgres, "why did this rank here?" is a query, not a black box:
SELECT id, summary, confidence
FROM pgmnemo.agent_lesson
ORDER BY confidence DESC;
Honest scope: on raw keyword recall, BM25 still beats us — and we publish that. The win isn't a higher recall@k; it's a memory layer with a feedback loop you can inspect in SQL.
— — —
Известная проблема RAG и памяти агентов: ретривал ранжирует по похожести, но похожесть — это не польза. Векторное хранилище будет раз за разом возвращать заметку, которая выглядит релевантной и каждый раз оказывалась неверной, — потому что ничто не связывает «что достали» с «помогло ли это». Обратной связи нет.
Команды латают это вручную: курируют память, чистят устаревшее, прописывают приоритеты руками. Не масштабируется и не проверяемо.
pgmnemo (расширение Postgres для памяти агентов) встраивает обратную связь прямо в базу. У каждой сохранённой записи — здесь их называют «уроками» — есть оценка confidence, которую читает ретривал. После задачи вы фиксируете исход одной строкой:
SELECT pgmnemo.reinforce(lesson_id, 'success'); -- или 'failure'
Доверие растёт у того, что помогло, и падает быстрее у того, что раз за разом предшествовало провалам. Запись с хорошей историей всплывает первой, балласт тонет. Без отдельного сервиса, без LLM на записи.
И раз это Postgres, «почему оно тут в выдаче?» — это запрос, а не чёрный ящик:
SELECT id, summary, confidence
FROM pgmnemo.agent_lesson
ORDER BY confidence DESC;
Честная граница: по сырому keyword-recall BM25 нас всё ещё обыгрывает — и мы это публикуем. Выигрыш не в более высоком recall@k, а в слое памяти с обратной связью, который можно проверить в SQL.
Teams patch this by hand: curate memories, prune stale ones, hard-code priorities. Doesn't scale, isn't auditable.
pgmnemo (a Postgres extension for agent memory) puts the feedback loop in the database. Each stored memory — it calls them "lessons" — has a confidence score the retriever reads. After a task, you record the outcome in one line:
SELECT pgmnemo.reinforce(lesson_id, 'success'); -- or 'failure'
Confidence rises for what helped, falls faster for what kept preceding failures. Good track records surface first; dead weight sinks. No extra service, no LLM on the write path.
And because it's Postgres, "why did this rank here?" is a query, not a black box:
SELECT id, summary, confidence
FROM pgmnemo.agent_lesson
ORDER BY confidence DESC;
Honest scope: on raw keyword recall, BM25 still beats us — and we publish that. The win isn't a higher recall@k; it's a memory layer with a feedback loop you can inspect in SQL.
— — —
Известная проблема RAG и памяти агентов: ретривал ранжирует по похожести, но похожесть — это не польза. Векторное хранилище будет раз за разом возвращать заметку, которая выглядит релевантной и каждый раз оказывалась неверной, — потому что ничто не связывает «что достали» с «помогло ли это». Обратной связи нет.
Команды латают это вручную: курируют память, чистят устаревшее, прописывают приоритеты руками. Не масштабируется и не проверяемо.
pgmnemo (расширение Postgres для памяти агентов) встраивает обратную связь прямо в базу. У каждой сохранённой записи — здесь их называют «уроками» — есть оценка confidence, которую читает ретривал. После задачи вы фиксируете исход одной строкой:
SELECT pgmnemo.reinforce(lesson_id, 'success'); -- или 'failure'
Доверие растёт у того, что помогло, и падает быстрее у того, что раз за разом предшествовало провалам. Запись с хорошей историей всплывает первой, балласт тонет. Без отдельного сервиса, без LLM на записи.
И раз это Postgres, «почему оно тут в выдаче?» — это запрос, а не чёрный ящик:
SELECT id, summary, confidence
FROM pgmnemo.agent_lesson
ORDER BY confidence DESC;
Честная граница: по сырому keyword-recall BM25 нас всё ещё обыгрывает — и мы это публикуем. Выигрыш не в более высоком recall@k, а в слое памяти с обратной связью, который можно проверить в SQL.
pgmnemo v0.8.0 — Token-Economy Navigation 🧭
Loading every retrieved chunk into an agent's context is what makes retrieval expensive. v0.8.0 splits recall into a cheap LOCATE and an on-demand EXPAND, so you spend context tokens only on what you actually use — ranked across vector + graph + JSONB in one SQL call, inside your own Postgres.
What's new:
•
•
• Pattern: locate within a token budget → expand only the rows you need.
•
•
Fixed (important):
Graph-proximity in
Upgrade:
Additive — new functions and one column with a default; brief lock only.
pg_regress: 21/21 ✅ (incl. a cold-start ranking guard)
Full changelog
Loading every retrieved chunk into an agent's context is what makes retrieval expensive. v0.8.0 splits recall into a cheap LOCATE and an on-demand EXPAND, so you spend context tokens only on what you actually use — ranked across vector + graph + JSONB in one SQL call, inside your own Postgres.
What's new:
•
navigate_locate(query_embedding, query_text, token_budget_chars, jsonb_filter) — returns only IDs + scores, stopping once the cumulative budget is reached. No content loaded. The JSONB predicate is pushed into the candidate scan (uses the existing GIN index).•
navigate_expand(ids, expand_fields, graph_expand_depth, ...) — fetches full text for the IDs you choose, plus optional causal/temporal graph neighbours.• Pattern: locate within a token budget → expand only the rows you need.
•
reembed() / reembed_batch() / recompute_content() — safe in-place updates that coexist with live ingestion (no graph rebuild).•
source_type column for origin classification.Fixed (important):
Graph-proximity in
recall_hybrid / navigate_locate was an additive term that could outweigh the retrieval signal by ~10×, burying a perfect vector+BM25 match beneath merely graph-connected rows — a cold-start problem for new or unconnected lessons. It is now a multiplicative tie-breaker: it only re-orders already-relevant candidates. Regression-guarded — a perfect match now reaches the top-3 with the graph term off and at the default weight, even against a dense connected cluster.Upgrade:
ALTER EXTENSION pgmnemo UPDATE TO '0.8.0';
Additive — new functions and one column with a default; brief lock only.
pg_regress: 21/21 ✅ (incl. a cold-start ranking guard)
Full changelog
🤝1
pgmnemo v0.8.1 — Adoption & Docs 📘
A docs-focused release: making it fast for an agent (or the human wiring one up) to understand what pgmnemo is and adopt it in one read. No schema change.
What's new:
•
• Positioning reframed across README / POSITIONING / WHY_PGMNEMO: pgmnemo ranks across vector + graph + JSONB in one SQL query plan inside your own Postgres — single-plan multimodal fusion, with the 0.8.0 token-economy
• "Disabling the provenance gate" FAQ in
• The gate's rejection message now names both
• Resolves adoption issues #18 (GUC access pattern), #19 (Docker install without a compiler), #20 (
Upgrade:
Body-only function refresh; no schema or data change.
pg_regress: 21/21 ✅
Full changelog
A docs-focused release: making it fast for an agent (or the human wiring one up) to understand what pgmnemo is and adopt it in one read. No schema change.
What's new:
•
AGENTS.md — a single canonical integration guide: what pgmnemo is, when it fits (and when it doesn't), every function with a minimal working SQL example, and copy-paste adoption recipes (agent memory loop, token-economy retrieval, multi-tenant scoping, incremental updates). Point an agent at it and it can self-assess and integrate.• Positioning reframed across README / POSITIONING / WHY_PGMNEMO: pgmnemo ranks across vector + graph + JSONB in one SQL query plan inside your own Postgres — single-plan multimodal fusion, with the 0.8.0 token-economy
navigate_locate → navigate_expand pattern as the headline.• "Disabling the provenance gate" FAQ in
SQL_REFERENCE.md — the gate requires commit_sha/artifact_hash on every write; the FAQ shows how to relax it (SET / SET LOCAL / ALTER DATABASE / ALTER ROLE pgmnemo.gate_strict = 'warn'|'off') — or, better, just pass provenance and keep it.• The gate's rejection message now names both
'warn' and 'off' and points at the FAQ (a recurring "how do I turn this off?" question).• Resolves adoption issues #18 (GUC access pattern), #19 (Docker install without a compiler), #20 (
stats() diagnostics), #24 (orphan recovery docs).Upgrade:
ALTER EXTENSION pgmnemo UPDATE TO '0.8.1';
Body-only function refresh; no schema or data change.
pg_regress: 21/21 ✅
Full changelog
Setup FAQ — two questions we keep getting from people installing pgmnemo.
1) Which embedding model? Is bge-m3 enough?
Yes. One gotcha: pgmnemo expects exactly 1024-dim vectors — it's in the schema, and ingest() rejects any other size. bge-m3 (1024-dim, multilingual) is the validated default; serve it from LM Studio's /v1/embeddings endpoint. pgmnemo doesn't embed for you — you generate the vector and pass it to ingest() and recall(). Rule: same model for writes and queries, or cosine compares apples to oranges.
2) How do I turn off the commit-hash check?
That's the provenance gate. It requires commit_sha OR artifact_hash on insert so every memory carries provenance.
SET pgmnemo.gate_strict = 'off'; -- disable
SET pgmnemo.gate_strict = 'warn'; -- allow, but log a warning
Default is 'enforce'. Cleaner than turning it off: just pass commit_sha or artifact_hash on insert.
— — —
Setup FAQ — два вопроса, которые чаще всего задают при установке pgmnemo.
1) Какая модель эмбеддингов? Хватит ли bge-m3?
Да. Один нюанс: pgmnemo ждёт ровно 1024-мерный вектор — это зашито в схему, и ingest() отклонит другую размерность. bge-m3 (1024-dim, мультиязычная) — проверенный дефолт; поднимай её в LM Studio через эндпоинт /v1/embeddings. pgmnemo сам не считает эмбеддинги — ты генеришь вектор и передаёшь его в ingest() и recall(). Правило: одна и та же модель на запись и на запрос, иначе косинус сравнивает несравнимое.
2) Как отключить проверку хеша коммита?
Это provenance-gate. Он требует commit_sha ИЛИ artifact_hash при вставке, чтобы у каждой памяти был провенанс.
SET pgmnemo.gate_strict = 'off'; -- выключить
SET pgmnemo.gate_strict = 'warn'; -- пропускать, но логировать
По умолчанию 'enforce'. Чище, чем выключать: просто передавай commit_sha или artifact_hash при вставке.
1) Which embedding model? Is bge-m3 enough?
Yes. One gotcha: pgmnemo expects exactly 1024-dim vectors — it's in the schema, and ingest() rejects any other size. bge-m3 (1024-dim, multilingual) is the validated default; serve it from LM Studio's /v1/embeddings endpoint. pgmnemo doesn't embed for you — you generate the vector and pass it to ingest() and recall(). Rule: same model for writes and queries, or cosine compares apples to oranges.
2) How do I turn off the commit-hash check?
That's the provenance gate. It requires commit_sha OR artifact_hash on insert so every memory carries provenance.
SET pgmnemo.gate_strict = 'off'; -- disable
SET pgmnemo.gate_strict = 'warn'; -- allow, but log a warning
Default is 'enforce'. Cleaner than turning it off: just pass commit_sha or artifact_hash on insert.
— — —
Setup FAQ — два вопроса, которые чаще всего задают при установке pgmnemo.
1) Какая модель эмбеддингов? Хватит ли bge-m3?
Да. Один нюанс: pgmnemo ждёт ровно 1024-мерный вектор — это зашито в схему, и ingest() отклонит другую размерность. bge-m3 (1024-dim, мультиязычная) — проверенный дефолт; поднимай её в LM Studio через эндпоинт /v1/embeddings. pgmnemo сам не считает эмбеддинги — ты генеришь вектор и передаёшь его в ingest() и recall(). Правило: одна и та же модель на запись и на запрос, иначе косинус сравнивает несравнимое.
2) Как отключить проверку хеша коммита?
Это provenance-gate. Он требует commit_sha ИЛИ artifact_hash при вставке, чтобы у каждой памяти был провенанс.
SET pgmnemo.gate_strict = 'off'; -- выключить
SET pgmnemo.gate_strict = 'warn'; -- пропускать, но логировать
По умолчанию 'enforce'. Чище, чем выключать: просто передавай commit_sha или artifact_hash при вставке.
pgmnemo v0.8.2 — adopter fixes + MCP that just works 🔌
A patch driven by real production reports. No schema change —
Fixed (the "my recall is empty" class):
• Consistent
• No more silent empty recall. If recall returns nothing while unprovenanced "ghost" lessons match, pgmnemo now emits a
• Docs:
Added — the MCP is now drop-in:
• Self-embedding via
• Docker image — run the MCP in a container so its deps don't fight your agent libraries on Linux:
Multi-arch (amd64/arm64), auto-built and pushed to Docker Hub on every release tag.
pg_regress: 22/22 ✅ · MCP ingest+recall verified end-to-end inside the published image.
🐳 Docker Hub · 📦 GitHub release · 📝 Full changelog
A patch driven by real production reports. No schema change —
ALTER EXTENSION pgmnemo UPDATE TO '0.8.2'.Fixed (the "my recall is empty" class):
• Consistent
include_unverified parsing — every recall path now accepts on/true/1 (one function only honoured the literal 'on').• No more silent empty recall. If recall returns nothing while unprovenanced "ghost" lessons match, pgmnemo now emits a
NOTICE telling you exactly why and how to include them — instead of leaving you guessing.• Docs:
ALTER DATABASE SET applies only to new connections (restart the MCP/pool).Added — the MCP is now drop-in:
• Self-embedding via
EMBEDDING_SERVER — point the MCP at any OpenAI-compatible embeddings endpoint and it embeds queries/lessons itself → real vector+BM25 hybrid recall, no out-of-band vectors. Unset → text-only fallback."env": { "DATABASE_URL": "...", "EMBEDDING_SERVER": "http://host:1234/v1/embeddings" }• Docker image — run the MCP in a container so its deps don't fight your agent libraries on Linux:
docker pull gaidabura/pgmnemo-mcp:0.8.2
Multi-arch (amd64/arm64), auto-built and pushed to Docker Hub on every release tag.
pg_regress: 22/22 ✅ · MCP ingest+recall verified end-to-end inside the published image.
🐳 Docker Hub · 📦 GitHub release · 📝 Full changelog
pgmnemo v0.8.3 — docs patch 📝
We ran a cold first-setup test (a fresh engineer adopting pgmnemo with zero insider knowledge) and fixed exactly what tripped them. No schema or scoring change — SQL is byte-identical to v0.8.2.
Fixed:
• The "Verify install" smoke SQL was broken — the very first command a new user runs. It failed twice (the bare
• MCP tool-argument contract was mis-documented — the README implied
• Version drift in install examples (pinned
Upgrade is a no-op at the SQL level:
🐳 Docker Hub · 📦 GitHub release · 📝 Full changelog
We ran a cold first-setup test (a fresh engineer adopting pgmnemo with zero insider knowledge) and fixed exactly what tripped them. No schema or scoring change — SQL is byte-identical to v0.8.2.
Fixed:
• The "Verify install" smoke SQL was broken — the very first command a new user runs. It failed twice (the bare
NULL couldn't resolve the ingest() overload, and the example lesson was under the 20-char minimum). Now correct: NULL::vector(1024), 3::smallint, valid-length text.• MCP tool-argument contract was mis-documented — the README implied
ingest took a nested metadata dict, but the real schema exposes text/role/topic/importance/project_id/commit_sha/artifact_hash/metadata as top-level args. An agent following the old docs would silently mis-scope its lessons. Fixed + defaults documented.• Version drift in install examples (pinned
v0.8.1) → current, plus a from-zero MCP quickstart (fresh Postgres → extension → MCP container → verify).Upgrade is a no-op at the SQL level:
ALTER EXTENSION pgmnemo UPDATE TO '0.8.3'
🐳 Docker Hub · 📦 GitHub release · 📝 Full changelog
pgmnemo v0.9.0 — the token-economy release ⚡
This release makes
Fixed — the moat that wasn't shipping:
• navigate_locate budget now counts what it returns. It was charging the full lesson text against your token budget while only returning a 50-char preview — so a 2000-char budget gave you ~8 candidate IDs instead of ~40. Locate is now genuinely cheap: rank, hand back many lightweight IDs within budget, expand only what you pick. ⚠️ Behavioral change: the same
Added:
•
• Typed-content columns —
• Selective embedding — a precise fact can be ingested with a NULL embedding (indexed by keyword/btree) without being excluded from recall.
Changed — recall_hybrid is fast on real corpora:
• Rewrote
Upgrade:
📦 GitHub release · 📝 Full changelog
This release makes
navigate_locate actually cheap, scopes it to a project, fixes hybrid recall on real-size corpora, and lays the schema groundwork for typed/multimodal content.Fixed — the moat that wasn't shipping:
• navigate_locate budget now counts what it returns. It was charging the full lesson text against your token budget while only returning a 50-char preview — so a 2000-char budget gave you ~8 candidate IDs instead of ~40. Locate is now genuinely cheap: rank, hand back many lightweight IDs within budget, expand only what you pick. ⚠️ Behavioral change: the same
token_budget_chars now yields ~5× more IDs — lower your budget proportionally to keep prior result counts.Added:
•
project_id_filter on navigate_locate — scope locate to one project/namespace (parity with recall_hybrid; backward-compatible, optional 5th arg).• Typed-content columns —
content_type, blob_ref, doc_ref (all nullable, additive) — foundation for typed and multimodal memory.• Selective embedding — a precise fact can be ingested with a NULL embedding (indexed by keyword/btree) without being excluded from recall.
Changed — recall_hybrid is fast on real corpora:
• Rewrote
recall_hybrid as two bounded retrieval arms (vector + keyword) fused by RRF, so HNSW is actually used instead of scanning every row. On a 2500-row corpus: identical top-10 (Jaccard 1.00, recall@10 78.8% vs 78.1%) at lower latency; on production-scale corpora it ends the multi-second full-scan that previously timed out.Upgrade:
ALTER EXTENSION pgmnemo UPDATE TO '0.9.0'
📦 GitHub release · 📝 Full changelog