Give Your Coding Agents a Memory You Own

Details

Coding agents start every session blind — the rationale behind past decisions evaporates when a session ends. Funes, released by Hugging Face engineer David Corvoysier, tackles this by building a persistent memory index from the agent traces already sitting on your machine. Installation is one curl command; adding memory to an agent takes a single line (funes add claude, with Codex, pi, and Hermes also supported). That command builds the initial index, installs automation to index each completed turn, and equips the agent with recall and get tools. The agent can then reach for past decisions autonomously — returning the original text with full provenance (agent, timestamp, session, and turn), not a summary. Under the hood, funes parses traces into turn-and-block chunks, embeds them with a pinned local model into a Lance dataset, then queries with combined vector and BM25 search, fuses rankings via a cross-encoder, and reweights results by recency. No external ML runtime is required — all embedding and reranking runs on-device. Memory can optionally sync to a private Hugging Face dataset the user owns, enabling cross-machine recall without vendor lock-in. The project is early-stage and the published article cuts off mid-sentence, but it represents a concrete, local-first approach to the growing problem of stateless coding agents discarding months of accumulated context.