Folklore: an open-source Letta (MemGPT) alternative
Folklore is an open-source (MIT), local-first memory layer for AI agents and an alternative to Letta. Unlike Letta — a database-backed agent server that needs an LLM — Folklore runs locally with no server, no database, and no API key, and adds signed provenance plus peer-to-peer federation.
Star Folklore on GitHub →Folklore vs Letta
| Dimension | Folklore | Letta (MemGPT) |
|---|---|---|
| Architecture | Local library / MCP server — nothing to host | Agent server + database (e.g. Postgres) |
| Footprint | CPU-only, no API key | Needs a running server + an LLM |
| Signed provenance / poison-defense | Yes — flip-ASR → 0 | No provenance field |
| Peer federation | Yes (P2P, shipping) | No — single-user |
| License / cost | Open source (MIT), free, self-host | Open source + hosted cloud |
| Retrieval (BEIR SciFact NDCG@10) | 0.7522, CPU — leads CPU retrievers | Not published as BEIR scores |
Figures from the public, reproducible benchmark; labeled measured vs simulated, negatives kept in.
When Letta may fit better
Letta is a full agent framework with self-editing memory and a managed server — if you want an opinionated agent runtime with built-in memory management and are happy running a server + LLM, it's a strong choice. Folklore is the better fit when you want a lightweight, local, poison- defensible memory layer that drops into any MCP harness, runs CPU-only with no server or key, and compounds across peers.
FAQ
Is there an open-source Letta alternative?
Yes — Folklore (MIT, local-first, CPU-only). No server/DB/LLM-key, plus signed provenance and peer federation. Repo + benchmark.
Do I need a server like Letta?
No — Letta runs an agent server + database; Folklore runs locally as a library / MCP server with neither.
Does Letta have provenance or federation?
No — Letta's memory is single-user, no signed provenance, no cross-user federation. Folklore adds both.
Last updated: 2026-06-20. Folklore works alone (local) today; P2P federation is shipping. Every number traces to the public benchmark, negatives included. See the full memory-tool comparison.