funes

funes gives coding agents a searchable memory of past work. It's for developers who want earlier decisions, reasoning and findings to remain available when they start another session or switch agents.

• Memory across agents. Search Claude Code, Codex, pi and Hermes conversations together, including subagent transcripts, with each result identifying its source agent.
• Local search. Embedding and reranking models run locally, combining meaning-based search with keyword matching over a local Lance dataset.
• Answers with session references. An agent can answer questions using retrieved history and name the sessions behind its answer; cited conversation turns are available in full.
• Automatic indexing. Claude Code, Codex and Hermes integrations keep the memory current as conversations progress, so later sessions can recall earlier work.
• Optional sharing. Publish memory datasets to the Hugging Face Hub for other machines or teammates to search, or read someone else's published memory.

funes is open source under Apache 2.0. It supplies retrieval models, but relies on your coding agent's generative model for answers. Through pi, that model can run locally or use the Hugging Face router; local retrieval doesn't make a cloud-backed agent local.

Publishing sends the memory dataset to the Hub. Repositories created by funes are private by default, while existing repositories keep their visibility. It redacts credentials during indexing and blocks publishing chunks that still contain detected secrets.