Give Your Coding Agents a Memory You Own

funes addresses a practical friction in AI-assisted coding: the loss of reasoning context when switching agents or machines. The tool converts agent session logs, already rich with search attempts, errors, and decisions, into a structured memory layer. This memory is queryable via recall, which returns original text with provenance, allowing engineers to revisit past rationale without manually reconstructing sessions. The design avoids cloud dependencies by default, embedding and reranking locally, which may appeal to teams with strict data governance policies.

The tool’s architecture is built around a deterministic pipeline that parses agent traces into a consistent format, chunks them, and embeds them using a pinned local model. Querying combines vector and BM25 search, with reranking and recency weighting to surface relevant context. This approach ensures that memory spans multiple agents (Claude Code, Codex, pi, Hermes) and retains raw evidence, avoiding the lossy distillation common in summarization-based systems. However, the reliance on local embedding models may introduce hardware constraints for large-scale deployments.

funes extends memory across machines by binding to a Hugging Face dataset, which acts as a shared, versioned repository. Publishing is incremental, with secrets redacted during indexing and again before upload. The Hub’s existing access controls apply, but the memory remains a dataset owned by the user, not a rented service. This model avoids vendor lock-in but requires engineers to manage dataset permissions and storage, which could complicate adoption for teams without existing Hugging Face workflows.

Integration is designed to be lightweight: a single command adds funes to an agent, automating indexing and recall. The tool’s local-first design means no account or cloud service is required for basic use, but cross-machine memory relies on Hugging Face’s infrastructure. While the security scanner for secrets is documented, its effectiveness depends on the patterns it recognizes, which may not cover all edge cases. Engineers will need to validate its redaction for their specific use cases, particularly in environments with strict compliance requirements.