Give Your Coding Agents a Memory You Own

Hugging Face Releases Funes to Give Coding Agents a Persistent Memory

Hugging Face has introduced funes, a new open-source tool designed to solve a persistent problem for developers using AI coding assistants: the lack of memory between sessions. The tool creates a durable, queryable memory layer from an agent's session logs, allowing different agents on different machines to access past reasoning, decisions, and abandoned approaches. This directly addresses the issue of agents starting every new task from a blank slate, enabling them to build on historical context without manual intervention from the user.

Local-First Indexing with Optional Cloud Sync

funes operates as a single binary that integrates with existing agents like Claude Code, Codex, pi, and Hermes. By default, it works entirely on the user's machine, parsing session traces, embedding content with a local model, and storing it in a local Lance dataset. The retrieval process uses a hybrid search model combining vector and BM25 search, which is then reranked with a cross-encoder to improve relevance. For developers working across multiple machines, funes can optionally bind this local memory to a private Hugging Face dataset, which keeps the history synchronized.

• Cross-Agent Memory: A single memory store can be read from and written to by multiple supported agents, ensuring continuity even when switching tools.
• Local-First Privacy: All indexing, embedding, and reranking operations run locally by default, without sending session data to a third-party service.
• Data Provenance: Every recalled piece of information is traced back to its exact source turn, including the specific agent, session, and timestamp.
• Secret Scanning: The tool includes a process to redact credentials during local indexing and performs a second scan to prevent secrets from being published to a shared repository.

Shifting from Agent Sessions to a Portable Asset

The introduction of a shared memory layer has significant implications for both team collaboration and open-source projects. A new team member's agent could query months of project decisions and technical dead-ends on day one. For open-source projects, maintainers can publish the development memory alongside a release, creating a searchable history of why the project evolved as it did. Performance benchmarks also suggest this approach is more efficient for long-running tasks, with tests showing that recalling specific context with funes was four to eight times cheaper than relying on an agent's internal context compaction or writing manual handoffs.

By decoupling an agent's reasoning history from the agent itself, funes repositions the most valuable asset as a portable, self-owned dataset, making the choice of a specific coding agent more flexible and less of a lock-in.