Hugging Face Releases 'funes', a Local Memory Layer for Coding Agents

Hugging Face has released funes, a durable memory layer designed to solve the problem of coding agents (such as Claude Code, Codex, pi, and Hermes) losing previous reasoning when sessions end. funes is a single binary that runs locally, performing embedding and reranking on the local machine, thus eliminating dependencies on accounts or hosted models.

Key Features and Architecture

• Preservation of Original Evidence: Returns original text and sources (agent, timestamp, session, turn) rather than summaries to prevent factual distortion.
• Hybrid Search: Combines vector search and BM25, reranking with a cross-encoder to improve accuracy.
• Inter-Agent Sharing: Records are stored in the same format across different agents, allowing them to share history.
• Security and Portability: Credentials are masked, and memory can be published as a private dataset via Hugging Face Hub or transferred to other devices.

Performance and Usage

Benchmark results show that funes' recall approach is more stable than compaction (summarization) for tasks requiring prior session knowledge, and is up to 8x cheaper than written handoffs. funes is released as open-source and utilizes local Lance datasets and Hugging Face Hub caching features.