huggingface/funes

What it solves

funes provides a durable, cross-agent memory system for AI coding agents. It prevents the loss of context, rationale, and decisions when switching between different AI tools (like Claude Code, Codex, pi, or Hermes) or starting new sessions, allowing agents to recall past work and findings spontaneously during a task.

How it works

The system indexes past agent sessions into a local Lance dataset using a deterministic pipeline: sources are parsed into turns and blocks, chunked, and embedded using pinned local models. It combines vector and BM25 search with reranking and recency weighting to retrieve relevant passages.

Users can "bind" a memory to an agent, which installs hooks to keep the index current every turn. These memories can be published as datasets to the Hugging Face Hub, enabling teammates or other machines to recall from the same knowledge base.

Who it’s for

Developers using multiple AI coding agents who want a persistent, shared memory of their development process, decisions, and technical rationale across different tools and models.

Highlights

• Cross-Agent Compatibility: Index sessions from multiple agents (Claude Code, Codex, pi, Hermes) into one memory.
• Hugging Face Integration: Publish memories as datasets to the Hub for easy sharing and collaboration.
• Automatic Indexing: Uses hooks to update memory per-turn and publish at session boundaries.
• Local Inference: Runs pinned local embedding and reranking models with no external ML runtime dependency by default.
• Grounded Answers: Provides an ask command to get answers grounded in the memory, citing specific sessions.