Give Your Coding Agents a Memory You Own

Hugging Face has released funes, an open-source memory layer designed to give coding agents a persistent, ownable memory across different machines and tools. According to the announcement, coding agents typically start tasks as strangers because reasoning and context from previous sessions disappear when they end. Although agents generate dense records of their work through traces, these logs usually require indexing and retrieval tools to be useful. Funes addresses this by parsing supported agent traces into a turn-and-block shape, chunking them, embedding them via a pinned local model, and writing them to a local Lance dataset. A single command installs the tool, building the first index and setting up automation to index completed turns incrementally. Agents can then independently use recall tools during conversations to reference past decisions, rationales, and findings, complete with original text and source provenance. Furthermore, users can bind their local memory to a Hugging Face dataset, allowing memories to follow them across different machines and enabling different agents, such as Claude Code and Codex, to share context. Before publishing to the Hub, funes scans data to redact credentials and secrets. Benchmarking suggests recall is cheaper and more effective than compaction or written handoffs, avoiding the loss of critical findings caused by summarization. Funes is available open source on GitHub.