Project MonetRequest demo
Home/Blog/Hugging Face Funes: Local Memory for Claude Code & Codex

AI · Project Monet Briefing

Hugging Face Funes: Durable Memory for Coding Agents

Funes is an Apache-2.0 local-first memory layer that indexes prior coding-agent sessions and makes their original evidence searchable across Claude Code, Codex, pi and Hermes.

Published 2026-09-04 · Updated 2026-09-04 · By Project Monet Editorial Team

Funes connecting multiple coding-agent sessions to a durable local memory index

01

What Hugging Face Funes is

Hugging Face released Funes on September 3, 2026 as a durable memory layer for coding agents. It is not a new model or agent: it indexes session traces already on your machine and lets supported agents retrieve past decisions, rationale and findings later.

The live repository lists Claude Code, Codex, pi and Hermes. All four normalize into one memory shape, so recall can cross agent boundaries while preserving the source agent, session, timestamp and turn.

02

How local recall works

Funes parses supported traces into a common turn-and-block format, chunks them, embeds them with a pinned local model and stores them in a local Lance dataset. Retrieval combines vector and BM25 search, reranking and recency weighting.

The agent gets recall and get tools. Recall returns original passages rather than only a distilled fact, while get opens the full turn and surrounding context. That keeps provenance available instead of making a lossy summary the only surviving record.

For the exact install and integration flow, continue to the Funes setup guide for Claude Code and Codex.

03

Local memory, Hub sharing and privacy

Local recall does not require a Hugging Face account. Embedding and reranking run on the user's machine, while the connected coding agent still performs the reasoning.

A memory can optionally be bound to a Hugging Face dataset repository. The current project documentation says repositories created by Funes are private by default; existing dataset repositories keep their current visibility.

04

Installation and current platform support

The current README lists prebuilt binaries for Linux x86_64, Linux aarch64 and macOS Apple Silicon. It does not list a Windows prebuilt binary, so native Windows support should not be assumed.

The default Linux build requires glibc 2.35 or newer according to the repository's build notes. Building from source requires Rust and protoc.

05

What the Funes benchmark does and does not prove

Hugging Face published a small project-authored handoff-versus-recall test and reports recall was eight times cheaper than a written handoff on one task and four times cheaper on another.

Those numbers are vendor/project-reported results from two narrow tasks, not an independent general benchmark of coding-agent memory systems. They are useful evidence about the intended design, not a universal cost-savings guarantee.

06

Where Funes is useful

  • Recovering why an architecture choice was rejected
  • Finding the root cause of a bug investigated in an older session
  • Moving work from Claude Code to Codex without losing investigation history
  • Sharing project memory across machines through a private dataset
  • Preserving technical dead ends and intermediate reasoning that never made it into stable docs

Funes complements READMEs, ADRs and project instructions rather than replacing them. Curated docs are better for stable rules; searchable traces are better for recovering the path that led to a decision.

07

Funes FAQ

Is Funes open source? Yes. The official repository is licensed under Apache License 2.0.

Does Funes replace Claude Code or Codex? No. It adds searchable memory; the connected agent's own model still performs the reasoning.

Does Funes support Windows? The current prebuilt-binary table does not list Windows. Recheck the live repository for later additions.

Sources

Primary and supporting sources

Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.

Project Monet

Useful signals. Clear decisions. Better digital work.

Project Monet turns relevant shifts in AI, creator tools and the web into practical context—and builds focused websites for businesses ready to grow.

Request a free homepage concept