pgmnemo

Overview

SQL-only extension requiring pgvector 0.7.0 or newer; upstream 0.15.0 and PIGSTY packages support PostgreSQL 17 and 18.

Version

Build

You can build the RPM / DEB packages for pgmnemo using pig build:

pig build pkg pgmnemo         # build RPM / DEB packages

Install

You can install pgmnemo directly. First, make sure the PGDG and PIGSTY repositories are added and enabled:

pig repo add pgsql -u          # Add repo and update cache

Install the extension using pig or apt/yum/dnf:

• Install
• pig
• dnf
• apt

pig install pgmnemo;          # Install for current active PG version

pig ext install -y pgmnemo -v 18  # PG 18
pig ext install -y pgmnemo -v 17  # PG 17

dnf install -y pgmnemo_18       # PG 18
dnf install -y pgmnemo_17       # PG 17

apt install -y postgresql-18-pgmnemo   # PG 18
apt install -y postgresql-17-pgmnemo   # PG 17

Create Extension:

CREATE EXTENSION pgmnemo CASCADE;  -- requires: vector

Usage

Sources:

• pgmnemo v0.14.2 README
• pgmnemo v0.14.2 usage guide
• pgmnemo v0.14.2 SQL reference
• pgmnemo v0.14.2 changelog
• pgmnemo v0.14.2 release notes
• pgmnemo v0.14.2 control file

pgmnemo stores agent memory in PostgreSQL and retrieves it through vector, BM25-style text, graph, metadata, temporal, provenance, and outcome-confidence signals. It installs into schema pgmnemo, requires the vector extension, and expects 1024-dimensional embeddings in its current SQL API.

Version 0.14.2 retains the 0.13 Bayesian confidence and outcome-use surface, and adds deterministic content classification, dry-run corpus reclassification, reversible near-duplicate consolidation, and a planner fix for hybrid HNSW recall.

Install

CREATE EXTENSION IF NOT EXISTS vector;
CREATE EXTENSION IF NOT EXISTS pgmnemo CASCADE;

SELECT pgmnemo.version();
SELECT * FROM pgmnemo.stats();

The v0.14.2 control file marks pgmnemo as trusted and non-superuser-installable when the required vector extension is available.

Ingest a Lesson

SELECT pgmnemo.ingest(
  p_role        := 'developer',
  p_project_id  := 1,
  p_topic       := 'security',
  p_lesson_text := 'Rotate signing keys after a compromise.',
  p_importance  := 4,
  p_embedding   := NULL,
  p_commit_sha  := 'abc1234',
  p_metadata    := '{"source":"incident-runbook"}'::jsonb
);

When pgmnemo.gate_strict is enforce, commit_sha or artifact_hash provenance is required. warn accepts an unverified write with an audit warning; off disables the gate.

Recall with Confidence Filtering

Hybrid recall combines embedding and text signals:

SELECT lesson_id, topic, score, match_confidence
FROM pgmnemo.recall_hybrid(
  '<1024-dimensional vector literal>'::vector(1024),
  'JWT rotation key compromise',
  10,
  'developer',
  1,
  0.4,
  0.4,
  60,
  'dag-2026-abc',
  ARRAY['note', 'fact'],
  0.40
);

The final p_min_score argument, added in 0.13.0, removes candidates whose match_confidence is below the threshold before LIMIT is applied. NULL preserves pre-0.13 behavior. The release notes suggest 0.40 as a starting point, not a universal value; calibrate it for the embedding model and feedback quality.

The same p_min_score concept is available in recall_fast, recall_lessons, and pooled recall entry points. recall_lessons routes to hybrid recall when both text and embedding are supplied and pgmnemo.disable_hybrid is off.

Record Outcomes

SELECT pgmnemo.reinforce(1001, 'success', true);
SELECT pgmnemo.reinforce(
  ARRAY[1001, 1002]::bigint[],
  'failure',
  false
);

The third p_used argument records whether the recalled memory was actually used. true or NULL increments use_count; false records the outcome without counting a use. Prefer an explicit value so analytics can distinguish ignored advice from used advice.

Under the default posterior mode, match confidence is:

(success_count + alpha)
/ (success_count + failure_count + alpha + beta)

The default Beta prior is alpha 1 and beta 1. Set pgmnemo.confidence_prior_alpha and pgmnemo.confidence_prior_beta between 0.01 and 100 when a different prior is justified.

Typed Memory and Navigation

Important write helpers include remember_fact, remember_event, remember_relation, add_edge, reembed, and recompute_content. remember_fact supersedes the active fact for an entity/property pair; events remain append-oriented; relations also populate the graph surface.

Use navigate_locate or navigate_locate_dispatch to select candidate IDs within a character budget, then navigate_expand_typed to fetch content and neighboring graph edges.

Configuration Index

• pgmnemo.confidence_mode: posterior by default; additive retains the legacy calculation.
• pgmnemo.confidence_prior_alpha and pgmnemo.confidence_prior_beta: Bayesian prior parameters.
• pgmnemo.confidence_boost_weight: contribution of confidence to ranking; defaults to 0, so confidence does not change rank unless enabled.
• pgmnemo.gate_strict and pgmnemo.include_unverified: provenance enforcement and retrieval.
• pgmnemo.disable_hybrid and pgmnemo.ef_search: recall strategy and HNSW search breadth.
• pgmnemo.track_recall_recency: whether recall updates last_recalled_at and recall_count.
• pgmnemo.max_query_text_chars, pgmnemo.tenant_id, and pgmnemo.test_project_floor: text, tenancy, and optional test-project controls.

The older confidence-delta settings are deprecated and ignored in posterior mode.

Caveats

Version 0.14 adds corpus-maintenance operations. They are read-only by default:

SELECT * FROM pgmnemo.reclassify_corpus();
SELECT * FROM pgmnemo.consolidate(
  p_similarity := 0.92,
  p_dry_run := true,
  p_role := NULL,
  p_limit := 100
);
SELECT * FROM pgmnemo.undo_consolidate(
  p_canonical_id := 42,
  p_dry_run := true
);

Set p_dry_run := false only after reviewing the result inside a transaction. In 0.14.2, reclassification touches only null or classifier-owned types and preserves curator-owned types such as event and relation. Consolidation marks noncanonical lessons superseded, writes edges, and accumulates evidence counts; undo_consolidate uses those edges to restore a selected cluster.

• Recall can write recency metadata. Disable pgmnemo.track_recall_recency for read-only analysis.
• The confidence model is only as reliable as reinforcement feedback. Avoid treating posterior values as calibrated probabilities without evaluation.
• HNSW, text, graph, and metadata indexes increase write and maintenance cost.
• The default confidence_boost_weight of 0 means p_min_score can filter results while confidence still contributes nothing to ranking.
• Classification is a deterministic keyword and regular-expression heuristic, not semantic review. Always inspect dry-run distributions and proposed duplicate clusters before applying corpus changes.