Hugging Face releases Funes, a local memory layer for coding agents

Funes indexes agent traces—records of codebase searches, approaches tried, errors encountered, documentation read, and direction changes—into a durable, queryable memory layer for coding agents such as Claude Code, Codex, pi, and Hermes.

The tool is designed to address the problem where agents start each new session as strangers to prior work, losing all reasoning from previous interactions. Funes builds memory from existing session logs, adding indexing, retrieval, ranking, and provenance so agents can recall past decisions and rationales during ongoing tasks.

Funes operates locally by default, using a single binary with no ML runtime dependency for its default inference backend. Embedding and reranking occur on the user’s machine, and the agent performs its own reasoning. Installation is via a shell command, and integration with an agent is a one-command process that builds the initial index and installs automation to index each completed turn.

Memory can optionally be published to a Hugging Face dataset owned by the user, private by default. The local memory is a Lance dataset, and the shared memory is a Hugging Face dataset. Before publishing, credentials are redacted during indexing, and a scanner documented in SECURITY.md removes potential secrets before anything reaches the Hub.

The tool supports recall and get commands: recall retrieves the original text with full provenance (agent, timestamp, session, turn), while get opens the full turn and surrounding context. The pipeline parses traces into a uniform turn-and-block shape, chunks them, embeds with a pinned local model, and writes to a local Lance dataset. Queries combine vector and BM25 search, fuse rankings, rerank with a cross-encoder, reweight by recency, and attach neighboring chunks.

Funes enables memory to follow work across machines by binding a memory to an agent, which publishes updates to a Hugging Face dataset at session boundaries. Running the same command on another machine allows the memory to travel, with remote datasets cached locally for fast queries. Ownership, access control, versioning, and distribution are handled via the Hugging Face Hub.