Notes from the commons.
Short essays on why agents keep paying for work that's already done, and what it looks like when knowledge compounds instead of evaporating.
Open LLM reasoning
Open weights freed the models; the reasoning they produce still evaporates. Treat reasoning traces as signed, shareable artifacts and the network stops paying twice — measured: 75% of repeat LLM calls avoided, zero false reuse.
2026-07-29 · 5 min The ProblemYour agent has amnesia
Every session, your coding agent rediscovers the same codebase, re-reads the same paper, re-derives the same answer, and you pay for it each time. Here's the bill nobody itemizes.
2026-06-15 · 4 min The CommonsStone soup for machines
In the folk tale, everyone drops one ingredient and the whole village eats. Federation is that, for inference: each peer contributes what it ground out, nobody pays twice.
2026-06-15 · 4 min Alone FirstNever research twice
The bar that buries most memory tools: be useful to one person, on day one, with no network at all. Why single-player value has to come before the commons.
2026-06-15 · 4 min ProvenanceSigned by a named hand
Anonymous answers of unknown trust are how RAG gets poisoned. Every record in Folklore is signed by a real, verified human, so a model can refuse what it can't trace.
2026-06-15 · 4 min