Postgres extension for agent memory: single-plan multimodal recall, token-budget navigation, zero-cost writes, optional provenance enforcement.

One CREATE EXTENSION command. Vector + BM25 + graph proximity + JSONB pushdown in one SQL query plan. Zero LLM inference per write. Token-economy navigation: locate IDs within a budget, expand content on demand. Provenance gate configurable via GUC: enforce / warn / off.

In-database agent memory substrate. Self-hosted. No new service. No vendor lock-in. EXPLAIN-able ranking.

pgmnemo serves three segments with one product, controlled by the gate_strict GUC:

Agents whose memory writes are traceable to independently-verifiable artifacts. Every ingest() call requires a commit_sha, document_hash, ticket_id, patient_record_id, or equivalent. Writes without provenance are rejected at the Postgres constraint layer — application code cannot bypass it.

| Segment | Typical artifact identifier | Compliance posture | |---|---|---|---| | RAG / document-grounded agents | document hash, chunk SHA, page revision ID | Optional (knowledge base audit) | | Customer support agents | ticket_id, conversation_id | Optional | | Clinical / healthcare AI | patient_record_id, clinical_note_version | Mandatory (HIPAA, GDPR) | | Legal AI (contract review, eDiscovery) | case_id, filing_id, citation_string | Mandatory (litigation hold, chain-of-custody) | | Software dev agents | commit_sha, pr_id | Optional (change tracking) | | Compliance / GRC AI | audit_event_id, control_id | Mandatory (SOC 2, ISO 27001, audit trail) |

Agents that build memory from multi-turn dialogue, sensor fusion, or ambient environment observation. No provenance artifact required. Set gate_strict = 'off' for unconstrained writes; set 'warn' for development audit logs without blocking writes.

Loading pre-existing memory, data migration, or legacy system bootstrap. Set gate_strict = 'warn' during the backfill, emit warnings for unverified rows, then reset to your production mode once backfill completes.

Agents need persistent memory in their control — not in a third-party API. pgmnemo puts memory where it belongs: inside your Postgres database.

Core value proposition:

1. In-database recall, zero new infrastructure. Vector search + BM25 hybrid scoring in pure SQL. No sidecar service, no managed vector DB, no API dependency. CREATE EXTENSION pgmnemo and you're done.

2. Zero LLM cost per write. Memory ingest is a SQL constraint check, not a model API call. Contrast with Mem0 ($0.17 per 1,000 writes for fact extraction) or Zep ($0.36 per 1,000 writes for contradiction resolution). pgmnemo's gate is compute-only.

3. Data stays in your Postgres. No data egress, no SaaS vendor lock-in, no cloud billing surprises. Control your own RLS, backup, encryption. HIPAA-aligned by architecture, not by policy.

4. Optional compliance gate for citation-grounded agents. When you need it, set gate_strict = 'enforce' — then every memory write is checked at the Postgres constraint layer before it commits. Hallucinated facts cannot silently accumulate. Unerasable audit trail in your database.

Who this is not designed for: If you want a fully managed SaaS product with pre-built agent integrations, Mem0 Cloud or Letta Cloud is the right choice. pgmnemo is for teams who want to own their agent infrastructure.

The differentiator claim: pgmnemo is the only Postgres extension that fuses vector (HNSW), BM25 full-text, graph-edge proximity, and JSONB metadata filtering into a single SQL query plan — with optional write-time provenance enforcement at the database constraint layer. The execution plan is inspectable via EXPLAIN, ranking is regression-testable with SQL, and no data leaves your database. This makes it simultaneously the simplest agent memory layer for conversational agents AND the only provenance-gated, EXPLAIN-able, token-economy-aware option for production agent systems.

Use pgmnemo if:

• Your Postgres is your primary datastore and you want memory in the same database (zero new service).
• You need to avoid per-write LLM costs (critical for high-velocity agents).
• You have compliance requirements (HIPAA, GDPR, litigation hold) — set gate_strict='enforce' for write-time provenance gates.
• You want single-plan multimodal recall (vectors + BM25 + graph proximity + JSONB pushdown in one SQL query plan), EXPLAIN-able and regression-testable.
• You need token-budget-aware retrieval — navigate_locate() + navigate_expand() let you control exactly how many characters your agent receives.
• You want outcome-learning feedback — reinforce() adjusts per-lesson confidence and match_confidence gives your agent an interpretable quality signal.
• You want data residency under your control (no vendor lock-in).

Use Mem0 if:

• You prefer a fully managed SaaS product with zero infrastructure overhead.
• You're OK with vendor lock-in and per-write LLM costs (~$0.17 per 1K writes).
• You want multi-agent cloud sync (shared memory across multiple agent instances).
• You want pre-built integrations (LangChain, LlamaIndex, CrewAI, etc.).

Use Zep/Graphiti if:

• You want structured knowledge-graph memory with rich edge semantics (semantic, temporal, causal, entity).
• You prefer self-hosted (Graphiti) with graph-native contradiction detection.
• You don't mind per-write LLM costs for contradiction resolution.

Use Letta if:

• You want an end-to-end agent framework, not just memory.
• Memory is one component of the agent, not your primary substrate.

Use Constructive AgenticDB if:

• You want pure vector memory in Postgres (no other frills).
• You don't need compliance gates or hybrid recall (BM25 + vectors).
• You prefer a minimal, vector-only approach.

Use GBrain if: your use case is a personal knowledge graph from Markdown files and you want zero-config Postgres (PGLite). Not for multi-agent fleet memory.

Use Memoir if: you want taxonomy-organized memory with Git-like versioning and deterministic path-based retrieval. Alpha-stage; no standard recall benchmarks yet.

Use agentmemory if: you want drop-in memory for a single coding agent (Claude Code, Cursor) with zero-config auto-capture hooks. Accept SQLite single-writer limitation and per-observation LLM cost.

Do not treat Odysseus as a memory competitor. It is a self-hosted AI workspace (ChatGPT alternative). Memory is a bolted-on ChromaDB session feature, not a substrate.

pgmnemo publishes numbers with confidence intervals and mandatory negative cells. Full protocol: docs/BENCHMARK_PROTOCOL.md.

We publish where we lose. A benchmark that shows only wins is indistinguishable from cherry-picking.

Apache 2.0 — github.com/pgmnemo/pgmnemo