Skip to content

memnos v0.1.0 — governed memory on PostgreSQL

Choose a tag to compare

View all tags
@thameema thameema released this 09 Jun 14:59
· 100 commits to master since this release
v0.1.0
21b25b1

First public release of memnos — self-hostable, governed, vendor-neutral long-term memory for AI agents, on a single PostgreSQL + pgvector. No second datastore, no LLM at query time (hybrid retrieval + a local cross-encoder reranker).

This is the PostgreSQL rearchitecture. The retired v1.x line (a different, graph-DB engine) has been removed; memnos is intentionally 0.x until adoption settles.

Install

pip install memnos        # server + admin/client CLI
pip install memnos-sdk    # lightweight REST client (LangChain / LangGraph / LlamaIndex adapters)

Then: memnos setup → memnos serve → open /admin. PostgreSQL (with pgvector) is a prerequisite. Full quickstart in the README.

What's in it

  • Governed by default — opaque bearer tokens, namespace ACL, audit log, usage/cost ledger, and an encrypted secret vault (AES-256-GCM) with ingest redaction.
  • One engine — pgvector HNSW + BM25 (tsvector) fused with RRF, then a local cross-encoder rerank; bi-temporal facts with deterministic single-valued supersession; offline consolidation into entity dossiers.
  • Agent integrations — one-command memnos claude-setup (MCP + auto recall/save hooks + /memnos), plus memnos agent-setup for Codex / Cursor / Windsurf / Claude Desktop. REST API + cross-platform CLI for everything else.
  • Reproducible benchmark — LoCoMo full-10 (1,542 QA) 58–61% under the gpt-4o judge, run from scratch with benchmarks/locomo_eval.py; all predictions published under benchmarks/results/.

Apache-2.0. Single-org self-host, local-first.

Assets 2
Loading