Hugging Face has released funes, a durable memory system that gives coding agents access to past reasoning and work across sessions. The tool addresses a fundamental problem where each new agent encounter treats a project as unfamiliar, with insights from previous work lost when sessions end. Funes automatically indexes agent traces—the detailed records of explorations, errors, and decisions already produced during work—making them searchable and retrievable for future tasks. The system works locally by default, with no ML runtime dependencies or external account required. Agents query their memory through recall and ask commands, with results linked directly to original turn context and metadata. Optional synchronization to Hugging Face datasets allows memory to follow users across machines, with access controls and automatic credential redaction built in. The implementation uses a combination of vector search, BM25 ranking, and cross-encoder reranking on session logs that are parsed into a unified turn-and-block structure. Funes supports multiple coding agents including Claude Code, Codex, pi, and Hermes, enabling unified memory across different tools. Installation requires a single binary download and one-command integration, with incremental indexing of new sessions rather than repeated full-history embedding. The approach positions agent memory as a personal dataset owned and controlled by the user rather than a subscription service.