Hugging Face released funes on September 3, an open-source memory layer that lets coding agents retain what happened in past sessions and recall it across different machines and different tools. The project, built by Hugging Face engineer David Corvoysier, targets a specific failure: every time a developer switches from Claude Code to Codex, or moves to a new laptop, the agent starts over, blind to decisions made a day or a month earlier.
funes works by indexing the trace files coding agents already leave behind on disk. According to Hugging Face, a single command called funes add builds a local index from an agent’s session history and installs a hook that indexes each completed turn going forward. Embedding and reranking run on the user’s machine using a pinned local model, so no account or hosted service is required for the tool to function day to day. Queries blend vector search with BM25 keyword matching, rerank results with a cross-encoder, and weight them by recency before handing back the original session text instead of a summary.
The cross-agent design is the project’s central claim. Hugging Face says the traces from Claude Code, Codex, pi and Hermes all land in a single shared shape, so a decision made in one agent can be retrieved by a different agent in a later session on a different host. A memory can also sync to a private Hugging Face dataset the user owns, which keeps the index current across machines and lets a team share the reasoning behind a project rather than just its finished code.
That portability is the detail worth isolating. Coding-agent vendors have every incentive to make their own session history sticky: a developer who has spent months accumulating context inside one tool faces a real switching cost the moment a competitor ships something better. funes treats a dataset the user owns as the storage layer, not an API rented back through a subscription, which turns memory portability into infrastructure rather than a feature. Whether a coding assistant is worth switching away from becomes a question about what you would lose, not only what the new tool can do, and funes is a bet that the honest answer should be nothing.
Hugging Face published a benchmark comparing funes recall against two alternatives: letting an agent compact its context and continue, or writing a manual handoff document before starting fresh. On two tasks whose answers required the original session’s reasoning, compaction succeeded on only one of the two; Hugging Face attributes the miss to detail the summarization step had already discarded. Recall beat both alternatives on cost, too: Hugging Face put it at roughly eight times below the price of a written handoff for the first task, and about four times below on the second.
AI Insiders covered a related argument two days ago: engineer Cal Paterson’s case for storing agent memory as plain Markdown files any tool can read. funes starts from a similar diagnosis, that agents already generate memory nobody retrieves, and answers it with an indexing and retrieval pipeline rather than a shared file format, trading Paterson’s simplicity for ranked recall across a much longer history.
funes is open source, and Hugging Face has published its own development history as a public memory dataset anyone can query with a --memory flag. Teams running a mixed stack, Claude Code on one project and Codex on another, should treat funes as the layer that decides whether that mix costs them anything in continuity over the next quarter.
Hugging Face published this project writeup, bylined to engineer David Corvoysier, on its blog on September 3, 2026.
