Funes gives coding agents a local memory that can follow you across tools and machines

Hugging Face released Funes on September 3 as an open-source durable-memory layer for coding agents including Claude Code, Codex, pi and Hermes. Funes ingests the session traces those agents already create, incrementally indexes them into a local Lance dataset and exposes retrieval designed to return both relevant passages and their provenance. Embedding and reranking run locally by default. A user can optionally bind the memory to a Hugging Face dataset they own, private by default, so the same memory can be reused on another machine or by another supported agent. Funes scans indexed content for credentials and performs another secret scan before publishing, but its own documentation does not claim those checks are infallible.

Coding agents are increasingly capable within a session but still routinely lose project decisions, failed approaches and architectural context when the session or tool changes. Funes makes that accumulated trace a portable artifact controlled by the developer rather than another vendor-specific conversation store. That can reduce repeated repository exploration and make switching agents less expensive, while keeping the default retrieval path local. The trade-off is governance: agent traces can contain credentials, internal code fragments and sensitive reasoning context, so turning them into a long-lived searchable dataset creates a new asset that needs access, retention and secret-handling rules.

Funes does not require developers to maintain a separate project notebook. It indexes existing agent sessions, preserving original text and provenance so a later retrieval can point back to the source turn rather than returning only an opaque synthesized memory.

The local memory is stored as a Lance dataset, while embedding and reranking use models that can run on the developer’s machine. That means a single-machine workflow does not require a separate hosted memory API or send every recall query to a remote service.

Funes integrations let the same indexed history be recalled from Claude Code, Codex, pi or Hermes. When a memory is bound to a Hugging Face dataset, another machine can download and cache it locally, allowing context to move with the user rather than staying trapped in one agent’s session database.

Funes redacts detected credentials while indexing and scans chunks again before publishing. The project documents the scanner and its limits rather than treating detection as a guarantee. Teams should therefore assume session traces may contain sensitive material and apply private dataset permissions, retention rules and independent secret scanning where the stakes justify it.