Hugging Face Funes Gives Coding Agents Portable Memory

Summary

On September 3, 2026, David Corvoysier introduced Funes, an open source single binary that turns existing Claude Code, Codex, pi and Hermes session traces into durable memory. `funes add` builds an initial index, adds `recall` and `get` tools, then indexes completed turns incrementally with bounded backfill. Recall returns original text, not summaries, with agent, timestamp, session and turn provenance; get opens surrounding context. A deterministic pipeline normalizes and chunks traces, embeds them with a pinned local model into append only Lance storage, combines vector and BM25 search, fuses rankings, reranks with a cross encoder, weights recency and attaches neighboring chunks. Embedding and reranking stay on device, the default backend needs no ML runtime dependency, and no account or Hub repository is required.

Binding memory with `funes add codex acme/funes-memory` publishes local state to a user owned Hugging Face dataset, private by default, updates it at session boundaries and lets another machine attach with the same command. Remote files cache locally for warm query speed, while the Hub provides ownership, access control, versioning and distribution. Credentials are redacted during indexing; publishing rescans every chunk and withholds suspected secrets, with coverage limits documented in SECURITY.md. `funes ask` queries local or shared memory read only, returns sourced answers, admits unsupported results and does not install an integration.

Shared memory crosses agents, models, hosts and teams, while public memories can expose release decisions, failed approaches and rationale through dataset cards and the Funes tag. In a two task benchmark requiring prior session knowledge, compaction succeeded once and never arrived once because its summary flattened key findings. Passage based recall was cheapest on both tasks, 8x cheaper than a written handoff on one and 4x on the other. Funes is available at github.com/huggingface/funes.