folklore

Folklore

A torrent swarm for agent reasoning.

Every peer shares the inference it already paid for, so your agent starts from what the swarm has worked out — not from zero. Like BitTorrent, but the payload is reasoning: debug traces, papers read, dead-ends ruled out. The more agents join, the less anyone re-derives. Local-first and useful alone on day one; compounding the moment a second node appears.

Site · Manifesto · Spec · RFC · Roadmap · npm

One session, two directions. You ask — the answer is pulled from @sam-rs's tree in 412ms, 0 web calls. Moments later the flow reverses — @tia-async pulls your trace, reputation ticking live in the menubar client. Like torrents, but the payload is reasoning. Reproduce it →

Web-search fallback rate, in simulator: 17% → 1% over 2,000 steps. Once anyone resolves a question, no one — you on Thursday, or any peer ever — pays for it again. § Proof

---

The problem

AI agents reason from zero, every single time. The same paper gets read again. The same dead-ends get walked again. The same conclusion gets re-derived — and re-billed — a thousand times across a thousand sessions.

You don't only pay Google. You pay OpenAI, Anthropic, and every paid endpoint, per token, to re-run inference over data someone already ground out yesterday. The work was done. The tokens were spent. Nobody kept the answer where the next agent could find it.

Memory tools exist — mem0, Letta, LangChain-style RAG — but they are single-user silos. They remember your chats. They don't gate the web, they don't carry provenance, and they certainly don't let another peer's hard-won debugging trace — the LLM inference they already paid for — become your starting point.

What Folklore is

A graph of your agent's reasoning that lives on your machine, answers before the web, and federates peer-to-peer.

Two bets, in order:

1. It works alone. Day one, zero peers, it already pays off — your own research, debugging, and grounded conclusions are cached and reused. You never start from zero again.
2. It compounds. Every peer running it works for the next: each one's past LLM inference — the reasoning already paid for in tokens — becomes your starting point, and yours becomes theirs. No teammate or org boundary; the whole network shares resolved inference so nobody re-spends tokens on an answer someone already ground out. The commons gets deeper the more peers draw from it.

Folklore is the name for knowledge that gets passed on — story to story, peer to peer — instead of relearned from scratch.

---

Day one: alone

No network required. Folklore sits between your agent and the web. Every research-shaped call — WebSearch, WebFetch, an arXiv pull, a fresh Read — is checked against your local graph first. If your graph already holds a confident answer, the call is satisfied from memory in milliseconds. If not, the call proceeds, and the result is auto-saved, signed by you, so the next time costs nothing.

This alone clears the bar that kills most memory projects: it is useful to one person, immediately.

folklore ask "how does mxbai-rerank compare to cross-encoder on long contexts?"
# → answered from your own graph if you've researched it; otherwise fetched + saved for next time

Retrieval quality is benchmarked, not asserted: 0.7522 NDCG@10 on full BEIR SciFact, CPU-only, 11 ms median — ahead of standard published BEIR SciFact retrievers (E5-base 0.737, E5-large 0.723, SPLADE 0.699, ColBERT 0.671, BM25 0.665); only a 3B GPU reranker (InRanker, 0.783) scores higher. No LLM grading an LLM.

Day N: together

Add a peer and your graphs become queryable to each other over libp2p. Now the first hop isn't even your graph — it's the question "what does the network already know about this?" Only when nobody holds the answer does anyone pay the web.

Every record carries a provenance chain: signed by its curator's cryptographic identity and verified GitHub handle, timestamped, traceable to the sources it grounded on. Not anonymous Stack-Overflow-grade trust — attributable, named, auditable knowledge.

The result, measured in simulation: web fallback collapses from 17% to 1% as the network warms.

Run folklore live and watch it happen: every line is a real fetch off your running node — someone else's agent reusing what yours already worked out. Reproduce it →

---

Philosophy

Folklore is a bet about how intelligence should be organized, not just a cache. The long version is the Manifesto; the short version:

Reasoning is labor. Every conclusion an agent reaches was paid for in tokens and electricity. An answer that dies with its context window is labor thrown into the sea — and every re-derivation of it is a tax, billed by the same few companies, for reasoning the world already owns.

Shared memory shrinks the model. The industry's answer to every gap is a bigger model — more parameters, more inference, more dependence on the center. There is another axis: remember more, infer less. An agent that starts from the network's resolved reasoning needs fewer tokens to reach further. Shared memory is the road to smaller, local, decentralized models — intelligence at the edge, not rented from a landlord.

No big brother in the loop. When one company hosts the memory, ranks the answers, and meters the reasoning, you are a tenant. Folklore has no landlord: your graph lives on your disk, your peers are chosen by you, every record is signed by a named hand instead of laundered through a platform. Local-first is sovereignty — if every peer but you vanished tonight, you'd lose nothing you made.

The commons compounds. A torrent gets healthier with every seeder; so does this. Every question resolved anywhere becomes a question nobody pays for again. Serving the network and serving yourself are the same motion.

---

Where it fits

Folklore is not a competitor to MCP or A2A. It's the missing third layer.

And against the memory tools you might compare it to: