Hugging Face Open-Sourced funes: Memory Your Coding Agent Owns, and Recall Measured 8x Cheaper Than Writing a Handoff

On 2026-09-03 Hugging Face published an open-source project called funes. It addresses something every AI-assisted developer has hit: every new session, the agent forgets everything you worked through yesterday. Verified 2026-09-08.

Their phrasing is apt: agents meet your project as a stranger each time.

1. What it does

funes turns agent session logs into retrievable memory:

1. Parses traces into standardised chunks
2. Embeds them locally on your machine, using pinned models
3. Stores them as Lance datasets
4. Retrieves with four signals combined: vector search + BM25 + cross-encoder reranking + recency weighting

The fourth point matters: this is not semantic search alone. Keywords, reranking and recency are all in the mix, because a question like “what did we decide last week” is not well served by embedding similarity on its own.

2. Supported agents, and installation

Officially listed: Claude Code, Codex, pi, Hermes.

Install:

curl -fsSL https://huggingface.co/buckets/huggingface/funes/resolve/install.sh | sh

Bind it to an agent:

funes add claude

(Swap claude for codex, pi or hermes.)

Two ways to use it:

• recall — the agent reaches past decisions automatically while working; you issue nothing.
• ask — you query the memory directly, without going through an agent.

⚠️ That install line pipes a remote script straight into your shell. It is the documented method, but opening the URL and reading it before running it is basic hygiene. That caution is ours, not Hugging Face’s.

3. The cost numbers they published

Hugging Face ran a handoff-vs-recall comparison — writing a handoff document versus recalling on demand. Their result: recall was 8x cheaper than a written handoff on one task, and 4x on the other.

The mechanism is straightforward. Handing over used to mean paying an agent to write a handoff document (tokens), then paying again to read the whole thing back in (more tokens). Retrieving a few relevant chunks when needed is naturally cheaper.

Note the limits: this is the authors’ own test, on two tasks, and we did not reproduce it. Treat 8x as a direction, not a guarantee.

4. Where the data lives

The question most people will ask:

On safety: credentials are redacted during indexing, with additional scanning to prevent secrets from being published.

Even so, agent session logs are among the more sensitive artifacts in a project — architecture, paths, and sometimes configuration people pasted in. Worth thinking about before turning sharing on.

5. Why it matters

The significance is less the feature than the location:

“The AI remembers you” has been a vendor-owned property — memory bundled into a product, tied to a subscription, reset when you switch tools. funes makes it a file you own: open source, on your machine, optionally synced to your own dataset.

In other words, memory is becoming portable. For anyone running two or three coding agents and unwilling to be locked into any of them, that is a usable step.

The project is on GitHub (github.com/huggingface/funes), built on existing open-source components (Lance, Hugging Face Hub).

How Claude Code’s weekly limits work: Claude Code weekly limit. Plan comparison on the Claude tool page.

Source read directly on 2026-09-08: Give Your Coding Agents a Memory You Own (Hugging Face, 2026-09-03). Supported agents, install commands, recall/ask, local Lance storage, credential redaction and the 8x/4x handoff-vs-recall figures are all as published there. We did not install or benchmark it. The curl | sh caution in section 2 and the view in section 5 are ours.