Give Your Coding Agents A Memory You Own

🔍 Read the full analysis: Give Your Coding Agents A Memory You Own on ThorstenMeyerAI.com

Hugging Face has introduced ‘funes’, a local-first memory layer designed to index and retrieve coding sessions from agents such as Claude Code, Codex, pi, and Hermes. This approach is detailed in the original analysis. This development addresses a persistent challenge in AI-assisted coding: the inability of agents to access previous decisions, reasoning, and technical details from earlier sessions, which hampers continuity and efficiency. The project enables developers to maintain ownership and control over their session data, supporting seamless transitions across projects and machines. For example, this open-source memory solution offers similar capabilities.

The funes system is distributed as a single binary that works with session records already stored locally on a user’s machine. When activated via the add command, it builds an index, installs recall and get tools, and automatically indexes completed turns from supported agents, including Claude Code and Codex. Updates to the index are incremental, allowing new activity to be added without reprocessing entire histories. The retrieval pipeline normalizes session traces into a common format, divides them into chunks, and stores them in a local Lance dataset. Searches combine vector similarity and BM25 ranking, with candidates reranked using a cross-encoder, and nearby chunks are attached to improve context. Retrieved passages include original session text, timestamp, agent, and session details, not summaries.

For cross-machine use, users can bind funes to a Hugging Face dataset, which publishes updates at session boundaries, enabling other agents or computers to access the same memory. Privacy controls are maintained as shared repositories are private by default, and data remains under user ownership, with no need for a Hugging Face account for local use. The design emphasizes ownership and provenance, allowing developers to verify the retrieved evidence instead of relying on the agent’s internal memory.

Implications for Developer Workflow and Session Continuity

The introduction of funes could significantly improve how developers manage AI-assisted coding. By enabling persistent, local memory, it allows for seamless continuation of projects, reducing repetitive exploration of codebases and enabling more informed decision-making. The ownership model enhances security and privacy, addressing concerns about cloud-based memory systems. If proven reliable, this approach could shift the paradigm from ephemeral session logs to durable, controllable memory, making AI tools more practical for long-term development efforts.

Existing Challenges in Coding-Agent Memory Management

Current AI coding agents typically operate on a session-by-session basis, with no built-in mechanism to recall previous interactions or decisions. Developers often manually copy-paste earlier conversations or lose valuable context when switching agents or devices. Prior efforts have focused on cloud-based memory solutions, but these raise security and privacy concerns. The idea that agent logs serve as a form of memory has been discussed, but without effective indexing or retrieval, these logs are often underutilized and difficult to leverage efficiently. funes builds on this understanding by offering a structured, indexed, and owner-controlled memory system that works locally and across machines.

“‘Funes treats memory as a local or user-controlled dataset, separating ownership from agent interpretation.'”

— Thorsten Meyer, project author

Unverified Claims and Security Considerations

The announcement does not include independent benchmarks for retrieval accuracy, latency, or security efficacy. It remains unclear how often irrelevant passages are retrieved, or if stale or outdated information surfaces after code updates. The security measures, such as secret scans during publishing, are documented but unverified, raising questions about privacy guarantees, especially for sensitive or regulated data. The effectiveness of cross-machine synchronization and how well the common trace format adapts to evolving session log structures are also uncertain.

Next Steps for Adoption and Validation

Developers and organizations will likely test funes in real projects to evaluate its reliability, security, and impact on workflow. Independent benchmarks and user reports will be critical to validate claims about retrieval quality and security. Future updates may focus on improving the robustness of cross-machine synchronization, handling of outdated information, and integration with more agent types. Monitoring how well the system scales with long-term, complex projects will determine its practical viability.

Key Questions

Can I use funes with my existing coding agents?

Yes, funes supports agents like Claude Code, Codex, pi, and Hermes, and works with session records already stored on your machine.

Is my session data secure with funes?

The system emphasizes ownership and privacy, with data stored locally and shared repositories being private by default. However, security claims have not been independently verified.

Can funes help me switch between different AI coding tools?

Yes, by providing a shared local memory, funes can facilitate switching between agents or machines without losing context, as long as the memory is synchronized.

Does funes require an internet connection?

No, funes operates locally, and remote synchronization is optional via Hugging Face datasets.

What are the limitations of funes at this stage?

Its retrieval accuracy, security effectiveness, and handling of evolving session data are still untested. Adoption in complex, long-term projects remains to be demonstrated.

Primary source: Hugging Face · via ThorstenMeyerAI.com