Give Your Coding Agents a Memory You Own

Hugging Face announced funes, an open-source durable memory layer for coding agents including Claude Code, Codex, pi, and Hermes. The tool is distributed as a single binary and is installed via a shell script hosted at huggingface.co, then attached to an agent with a command such as "funes add claude".

According to the announcement, funes builds its first index from sessions already on the machine, gives the agent recall and get tools, and installs automation that indexes each completed turn. Indexing is described as incremental, with new runs adding new turns rather than re-embedding the full history; older content can backfill in bounded steps. The recall function returns original text rather than a summary and identifies the agent, timestamp, session, and turn behind each result.

The described pipeline parses supported traces into a common turn-and-block shape, chunks them, embeds them with a pinned local model, and writes to a local Lance dataset. Queries combine vector and BM25 search, fuse rankings, rerank candidates with a cross-encoder, reweight by recency, and attach neighboring chunks. The default inference backend has no ML runtime dependency, and embedding and reranking run on the user's machine.

A memory can be bound to a Hugging Face dataset the user owns, private by default, using a command like "funes add codex acme/funes-memory". The announcement states credentials are redacted during indexing and that publishing scans every chunk again, withholding anything that still looks like a secret, with details documented in SECURITY.md. Remote memories are cached locally for warm queries.

funes also offers "funes ask", a read-only, one-question command that reads local memory by default. A published memory of funes development is available at huggingface/funes-memory. The announcement cites a handoff-vs-recall benchmark on two tasks and reports recall was 8x cheaper than a written handoff on one and 4x on the other, and that compaction arrived on one task but not the other. funes is at github.com/huggingface/funes.

In a linked comment, a maintainer of a separate tool, deja-vu, reported running funes 1.3.0 on 19,195 coding-agent sessions, with a full index taking 2h03m on an M4 Pro laptop (308k chunks) and 6.3s per query, and stated that a 30-day recency half-life halves rank-1 results on history older than a month. Hugo Laurin, VP of Machine Learning at Hugging Face, wrote in a comment that longmemeval does not need semantic search because BM25 gives the same ratio, and said funes shines when information is genuinely difficult to find.