funes gives coding agents a memory layer you host yourself

Coding agents meet your projects as strangers, the post argues: the reasoning from last Tuesday disappears when the session ends, and each new agent on each new host starts from zero. funes, described in a post dated September 3, 2026 by David Corvoysier, is presented as the answer to that problem. It is a durable memory layer for agents including Claude Code, Codex, pi and Hermes, and it is built from sessions already sitting on your machine.

The distinction the post draws is between an archive and memory. Session logs are “only potential memory”, and you cannot grep ten thousand turns to find out why a streaming parser was abandoned. funes parses supported traces into one turn-and-block shape, chunks them, embeds them with a pinned local model and writes them to a local Lance dataset. A query combines vector and BM25 search, fuses their rankings, reranks the candidates with a cross-encoder, reweights by recency and attaches neighbouring chunks. recall returns the original text rather than a summary and names the agent, timestamp, session and turn; a get command opens the full turn and its surrounding context.

Access is deliberately light. funes is a single binary, its default inference backend has no ML runtime dependency, and embedding and reranking happen on your machine. Local recall needs no account and no Hub repository. A single add command builds the first index, gives the agent recall and get tools, and installs the automation that indexes each completed turn; indexing is incremental, with older content backfilling in bounded steps.

Ownership is the second half of the pitch. Binding a dataset with the add command publishes the current memory and keeps it current at session boundaries, so the same command on another machine brings the memory along. The local memory is a Lance dataset; the shared one is a Hugging Face dataset, private by default. Credentials are redacted during indexing, and publishing scans every chunk again and withholds anything that still looks like a secret, a scanner documented in SECURITY.md. Remote reads are cached locally. A separate ask command is the read-only, one-question sibling: it reads local memory by default, does not install an integration or change the agent's persistent setup, and says so when the retrieved passages do not support an answer.

What the post does not supply is independent measurement, either of recall quality across long histories or of what the secret scanner misses; it points readers to SECURITY.md for the scanner's stated coverage. The bounded next step is small: add funes to one agent on one machine, work a normal session, then put a past decision to ask and check whether the provenance it cites is the turn you actually meant.