Hugging Face open-sourced Funes, a local memory layer for coding agents

Hugging Face has released Funes, an open-source memory layer for coding agents that indexes past sessions and makes them searchable to future ones. The project's own benchmark, published alongside the blog post on 3 September, reports that calling recall from a fresh session was around 8 times cheaper than pasting a written handoff on one investigation task and around 4 times cheaper on a second.

What Funes does

Funes indexes the session traces that Claude Code, Codex, pi and Hermes already write to disk, and exposes two tools, recall and get, that a running agent can call by itself. When an agent reaches a question it has looked into before, it can pull the relevant passages and their source from the index rather than starting the reasoning from the beginning. Hugging Face describes it in the blog post as durable memory for agents, licensed Apache 2.0, and available as a native binary with no ML runtime required. Memories can stay local, or be shared across machines by publishing them as private Hugging Face datasets, and the tool removes credentials before anything leaves the machine.

The benchmark behind the claim

The 8x and 4x numbers come from a benchmark Hugging Face published as a separate dataset. It compares five ways to move findings from one agent session to another: no transfer at all, a written handoff summary, Funes recall, a full session fork, and the /compact summary the coding agents already include. The metric is weighted API token cost, with cache reads and writes priced separately from fresh input and from output.

On the first task, described as triaging six routes to closing a 2x rerank performance gap against torch on Apple silicon, recall came in at roughly one eighth the cost of the written handoff, the fork and the compact summary. On the second, judging four candidate recall improvements against a pipeline ceiling, recall was about a quarter of the cost. Fresh sessions with no transfer never succeeded at either task. The benchmark is small, 41 rows and under a megabyte of results, so treat the numbers as an internal measurement of a specific technique rather than a general claim.

Working out again what the last session already knew

Anyone running an agent on a long project has watched the same investigations happen twice. A new session opens, reads the current state of the code, and works out again something the previous session had already figured out, because the compact summary either dropped the finding or paraphrased it into something the next agent cannot use. Funes is a wager that the cheapest fix to that pattern is not a longer summary or a bigger context window: it is a searchable index of the raw passages, keyed on what an agent actually asks about at recall time.

The design choices that make it interesting for a team are the local-first index and the credential redaction. A team can publish a shared memory as a private Hugging Face dataset, which keeps memory available across machines without the code or credentials leaving the organisation. It is early, the repository star count is still in the tens, and one benchmark on one team is not a general finding. But the reported cost gap is large enough that a group already spending real money on agent runs should reproduce it before the next long project starts.

Source

• Give Your Coding Agents a Memory You Own on the Hugging Face blog, by David Corvoysier
• funes-handoff-recall-benchmark dataset on Hugging Face
• huggingface/funes on GitHub