Folklore is a local-first, open-source memory and research layer for AI agents. Before an agent hits the web it checks a local knowledge graph; a confident hit is served from memory, a miss is fetched and saved with signed provenance so the agent never researches the same thing twice. It is CPU-only, needs no API key, works in any MCP harness, and is designed to compound peer-to-peer.

How does Folklore compare to mem0?

mem0 is a single-user, LLM-mediated memory store: it runs an LLM extraction on every write and carries no signed provenance and no peer federation. Folklore needs no LLM call per write, adds signed provenance (so poisoned context can be ranked down) and peer federation (so a team or network compounds). On a matched poison test a similarity-only store like mem0 is fully flipped by a crafted document (flip-ASR 0.625 toy, 1.0 at corpus scale) while Folklore's provenance ranking drives that to 0.

How does Folklore compare to Letta (MemGPT)?

Letta is a database-backed agent server with LLM-mediated self-editing memory — single-user, no signed provenance, no peer federation, and it needs a running server plus an LLM. Folklore is a local, CPU-only library/MCP server with provenance and federation. They overlap on 'agent remembers things'; Folklore's differentiators are attributable provenance and cross-peer compounding.

How does Folklore compare to LangChain RAG or Pinecone?

LangChain RAG and Pinecone-backed RAG are retrieval over a vector store. They do not gate the web, carry no signed provenance, and do not federate across users. Folklore measures at parity with a plain vector cache on single-user retrieval (it leads the CPU pack on BEIR SciFact at 0.7522 NDCG@10) and adds web-gating, provenance and federation on top.

Does Folklore beat GPU rerankers on retrieval?

No — and it does not claim to. On full BEIR SciFact, Folklore scores 0.7522 NDCG@10 CPU-only, ahead of the standard CPU retrievers (E5 0.737, SPLADE 0.699, ColBERT 0.671, BM25 0.665), but a 3B-parameter GPU reranker (InRanker, 0.783) scores higher. The public benchmark labels every number measured vs simulated and keeps the negatives.

Is Folklore free and open source?

Yes. Folklore is open source (MIT), CPU-only, and needs no API key. Code and the full benchmark are at https://github.com/usefolklore/folklore.

← Folklore

Folklore vs mem0, Letta & LangChain RAG — honest comparison

Most AI-agent memory tools are single-user stores ranked by embedding similarity. Folklore adds the two things they structurally lack: signed provenance and peer federation. Here's the benchmarked picture — including where Folklore is only at parity, and where it doesn't win at all.

Star Folklore on GitHub →

The short answer

On raw single-user retrieval, Folklore is at parity with a good vector cache (and leads the CPU field on BEIR SciFact). On provenance and federation, the other tools can't compete — not by a small margin, but because they have no provenance field and no cross-user sharing at all. So the honest pitch isn't "Folklore retrieves better"; it's "Folklore is the only one that's poison-defensible and compounds across peers."

Capability + benchmark matrix

Axis	Folklore	mem0 / Letta / LangChain RAG / Pinecone
Retrieval (BEIR SciFact NDCG@10)	0.7522, CPU-only — leads the CPU pack	parity at the retrieval layer (same embedder); not published as BEIR scores
Signed provenance / poison-defense	flip-ASR → 0.0	flip-ASR 0.625 (toy) → 1.0 (real corpus) — no provenance field
Federated compounding	~0.97 hit-rate @ 64 peers	flat ~0.31 — single-user, can't share
Footprint	CPU-only, no API key, any MCP harness	mem0/Letta need an LLM per write; Zep/Letta need a server
Beats a 3B GPU reranker?	No (InRanker 0.783 > 0.7522) — stated honestly	No

All figures from the public, reproducible benchmark; labeled measured vs simulated, negatives kept in.

Frequently asked

What is Folklore?

A local-first, open-source memory + research layer for AI agents. It answers from a local knowledge graph before the web and saves what it learns with signed provenance, so your agent never researches the same thing twice. CPU-only, no API key, any MCP harness, designed to compound peer-to-peer.

How does it compare to mem0?

mem0 is a single-user, LLM-mediated store (an LLM extraction per write) with no signed provenance and no federation. Folklore needs no per-write LLM call and adds provenance + federation. Under a matched poison test mem0-style similarity ranking is fully flipped; Folklore's provenance ranking takes the flip rate to 0.

How does it compare to Letta?

Letta is a DB-backed agent server with LLM self-editing memory — single-user, server + LLM required, no signed provenance, no federation. Folklore is a local CPU-only library/MCP server whose differentiators are attributable provenance and cross-peer compounding.

How does it compare to LangChain RAG / Pinecone?

Those are retrieval over a vector store: no web-gating, no provenance, no federation. Folklore is at parity on single-user retrieval (leads the CPU pack on SciFact) and adds those three on top.

Does it beat GPU rerankers?

No, and it doesn't claim to. CPU-only 0.7522 leads the CPU retrievers; a 3B GPU reranker (InRanker, 0.783) is higher. The benchmark keeps that negative visible.

Is it free / open source?

Yes — MIT, CPU-only, no API key. Code + full benchmark: github.com/usefolklore/folklore.

Looking for a specific alternative?

Dedicated head-to-heads: Folklore as a mem0 alternative · Folklore as a Letta (MemGPT) alternative.

See the code + full benchmark →

Last updated: 2026-06-20. Folklore works alone (local) today; peer-to-peer federation is shipping. Every comparison number traces to the public, reproducible benchmark, which labels measured vs simulated results and includes the cases where Folklore does not win. Whitepaper.