Funes is Hugging Face's durable memory layer for AI coding agents: it indexes past sessions from Claude Code, Codex, pi and Hermes into a single local dataset, and hands agents recall and get tools so they can pull prior decisions, rationale and findings into the current task. Built in Rust on a Lance dataset, Apache-2.0 licensed. Repository
The pipeline is source-agnostic: each agent is a TraceSource implementation that parses transcripts into a generic turn and block shape, which is chunked, embedded with a pinned local model and written to Lance. Recall fuses vector and BM25 search, reranks and reweights by recency, and every hit names the agent and session it came from, so a task started in Claude Code is recallable from Codex or Hermes later. The embedding model is stamped into the memory and queries with a different model are refused, and subagent transcripts are indexed too. The memory is a disposable derived artifact: the raw text stays in every row, so changing the embedder means a rebuild, not data loss. Indexing docs TraceSource trait
Memory publishes like a dataset. funes push uploads a local memory to a Hugging Face dataset repo owned by your account or org, private by default; a teammate, another machine or a published third-party memory is then readable with one flag, no indexing required. funes ask borrows an agent for a single grounded answer over a memory, naming the sessions it drew from. Publishing is guarded twice over: credentials are redacted at index time and an always-on gate refuses to push any chunk that still contains a secret. The installer verifies tagged release checksums against a manifest in the release bucket. Publishing docs Ask docs Installer
The project is young and moving fast: 291 stars, 22 forks and 551 commits on the main repo, with a rationale document explaining the design choices and why it differs from other memory tools. Inference runs on a built-in backend (Accelerate on macOS, pure Rust on Linux) so the default build has no ML runtime dependency; an ONNX Runtime backend is opt-in. Rationale