Provena: an evidence-traceable memory layer for LLM agents, open-sourced

No-Bicycle-4804 · reddit · 2026-09-29

The author open-sourced Provena, a provenance-focused memory layer for cross-session LLM agents. It stores source events separately from structured claims, links claims to evidence, and tracks scope, review status, validity, and source authority. Retrieval uses pgvector, but trustworthiness is decoupled from similarity—you can trace any claim back to its source and review history. It connects to Codex, Claude Code, Gemini CLI and other MCP clients, uses PostgreSQL as the record of record, and can run fact extraction locally with Ollama. Available on PyPI and GitHub; the author seeks feedback on handling conflicting or stale memories.

Related event: Provena Open-Sources Evidence-Traceable Agent Memory(2 posts)→

Original post →

More from coding & agent

coding & agent channel →