Hugging Face Announces funes, a Memory Layer for Coding Agents

Coding agents have faced the challenge of losing past inference content once a session ends. funes addresses this by indexing, searching, and ranking existing session logs on the machine, allowing agents to autonomously recall past decisions and the reasoning behind them.

The tool consists of a single binary. Since embedding and reranking processes are performed locally on the user's machine, there is no dependency on an ML runtime. Installation is completed with a single command, and it can be added to agents such as Claude Code, Codex, pi, and Hermes.

When an agent recalls its own memory, it displays the session name that served as the basis for the answer. This eliminates the need for users to manually paste past context.

Additionally, funes includes a feature to publish memory as a Hugging Face dataset and link it to user-owned datasets. This allows memory to be carried over even when working on a different machine. To protect privacy, a scanner that detects and removes sensitive information has also been implemented.