Hugging Face open-sources funes, a local-first memory layer for coding agents

Coding agents meet every project as a stranger: the reasoning from "last Tuesday" disappears when the session ends, and each new agent on each new host starts from zero. That is the problem funes, an open-source tool published by Hugging Face engineer David Corvoysier on September 3, sets out to solve. It is a durable memory layer for coding agents including Claude Code, Codex, pi, and Hermes, built from the session traces already sitting on your machine.

funes is a single binary whose default inference backend has no ML runtime dependency, with embedding and reranking happening on-device. After installation, one command such as "funes add claude" builds the first index, gives the agent recall and get tools, and installs automation that indexes each completed turn. Indexing is incremental: new runs add new turns instead of re-embedding the whole history, while older, deeper content backfills in bounded steps.

The most distinctive part is how memory is retrieved: with funes added, recall happens inside the conversation, and the agent reaches for its memory on its own, naming the session behind its answer. Recall returns the original text, not a summary, and shows exactly where it came from — the agent, timestamp, session, and turn. Each result includes a get command that opens the full turn and its surrounding context. Underneath, one deterministic pipeline parses every supported trace into the same turn-and-block shape, chunks it, embeds it with a pinned local model, and writes it to a local Lance dataset; queries combine vector and BM25 search, fuse their rankings, rerank with a cross-encoder, reweight by recency, and attach neighboring chunks.

The design yields three important properties: one memory across agents, since Claude Code, Codex, pi, and Hermes all write to the same shape and every hit says which agent produced it; raw evidence stays intact, with nothing distilled into a fact at write time and every result traceable to its source turn; and recall is local by default, requiring no account or Hub repository, with no hosted model processing your sessions and reasoning left to your coding agent.

funes is built on the idea that a memory is a dataset, not a service. Binding one is as simple as "funes add codex acme/funes-memory": the current memory is published and kept current, and running the same command on another machine carries the memory there. Under the hood, local memory is a Lance dataset while shared memory is a Hugging Face dataset you own, private by default. Credentials are redacted during indexing, and a second scan before publishing withholds anything that still looks like a secret, as documented in SECURITY.md. Remote memory files are cached locally so warm queries return at local speed, and users never rent memory back through a separate service's API.

A companion command, funes ask, is a read-only, one-question tool: it recalls relevant passages, hands them to a coding agent, and returns a grounded answer that names its sources, without installing an integration or changing the agent's persistent setup; retrieval misses are not papered over. Because a shared memory is not tied to the agent or model that created it, you can start a task in Claude Code and continue it in Codex next week, with the second agent recalling the first one's reasoning. The post outlines three scopes where this matters: across your machines, across a team where a new teammate's agent can retrieve months of decisions and dead ends on day one, and alongside open-source projects where maintainers can publish the sessions behind a release as a searchable CLAUDE.md.

On the accompanying handoff-vs-recall benchmark, funes was measured against the two usual ways out of an overlong session: letting the agent compact and carry on, or writing a handoff and starting fresh. Compaction, the default for most agents, arrived on only one of the two tasks because its summary flattened the findings that mattered; recall returns the passages themselves, so a finding does not have to survive summarization. Recall was also the cheapest of the three on both tasks, 8x cheaper than a written handoff on one and 4x on the other.

The tool takes its name from Borges's Funes the Memorious, the fictional man who could forget nothing — the opposite of the forgetting that "thinking" requires. funes is open source at github.com/huggingface/funes, leaning on Lance's append-only datasets, embedding models good enough to run locally, and the Hub's caching and content deduplication. The release also signals Hugging Face extending its platform beyond open weights and datasets toward "open working memory." What to watch: whether the broader agent ecosystem adopts it, how the supported-agent list grows, and whether the dataset-not-service framing reshapes long-horizon coding workflows.