This Tool Gives Your Coding AI a Memory That Lasts Between Sessions
A new open-source tool called funes lets AI coding assistants remember past decisions, failed experiments, and the reasoning behind your code, even after you close the session and reopen it days later.
Key points
- Funes is a free, open-source memory layer for AI coding assistants including Claude Code, Codex, pi, and Hermes, announced via Hugging Face.
- All indexing and search runs locally on your own machine; no cloud service processes your code or sessions.
- A shared memory option lets teams share months of past decisions with a new teammate on their very first day.
- In a small benchmark, recall via funes cost roughly eight times less than the leading alternative for retrieving prior session knowledge.
- Shared memories publish as private Hugging Face datasets by default, with credentials scrubbed before anything leaves your machine.
Every time you start a new session with an AI coding assistant, it meets your project as a stranger. Ask it why the team moved off the streaming parser three weeks ago, and it has no idea. That history vanished when the last session ended.
Funes, announced through Hugging Face, tries to fix that. It sits alongside the AI coding tools you already use and builds a searchable record of every decision, dead end and piece of reasoning those tools have produced. We've been tracking the AI memory space since 6 August 2026, and funes is the first tool we've seen that keeps that record entirely on your own hardware.
How does it actually work?
Funes watches your AI coding sessions and files them in a searchable index on your own computer. When a new session touches a topic covered in an old one, the AI can reach back into that index and retrieve the relevant passage, complete with the agent name, timestamp and session turn it came from.
Setup takes two commands. You install funes with a script, then attach it to your chosen coding tool. After that, you work normally. Indexing runs in the background, adding new turns incrementally rather than re-embedding your full history each time.
The retrieval system combines vector search (which finds passages conceptually similar to your question) and BM25 search (a keyword-matching approach familiar from standard search engines). A second model called a cross-encoder re-ranks the results to surface the most relevant ones. None of this touches an external server.
What does it return when you ask it something?
The full original text, not a summary. That distinction matters. When an AI compacts a long session into a shorter summary, it inevitably drops detail. Funes returns the actual passage, so nothing has to survive a summarisation step.
You can also query a memory directly from the command line: funes ask claude "what did we decide about the streaming parser". The tool retrieves relevant passages, hands them to the AI, and returns an answer that names its sources. If the passages don't support an answer, it says so rather than inventing one.
A small benchmark comparing three approaches to long-session recall found funes cost roughly eight times less than the next option and was the only approach that succeeded on both test tasks. Compaction, the method most AI agents use by default, failed one of the two tasks because its summary had dropped the relevant findings.
Can a whole team share the same memory?
Yes, and this is where funes gets interesting for organisations. Bind a funes memory to a Hugging Face dataset you own (private by default) and it follows you across machines. Any teammate who connects to that dataset can search months of past decisions, including rationale that never made it into a pull request or a document. For teams handling the kind of agent-to-agent handoffs our 3 September story on AI coding constraints described, a shared memory that persists across tools could matter more than people currently expect.
Before anything leaves your machine, funes scrubs credentials from the data, then scans again during the publishing step. The project documents what the scanner covers and where its limits are.
Open-source projects can also publish a memory alongside a release, giving contributors a searchable record of why the project works the way it does.
Common questions
Does funes send my code to any outside server?
No. Indexing, embedding (converting text into a searchable numerical form) and re-ranking all run locally. A shared memory publishes to a Hugging Face dataset you own and control, not to a third-party service.
Which AI coding tools does it support?
Funes currently supports Claude Code, Codex, pi and Hermes. The tool is open source, so support for other agents may follow.
What happens if funes cannot find a relevant answer?
It says so. The tool doesn't invent an answer when the retrieved passages don't support one. You can rephrase your question or let the AI agent search the memory more deeply during a normal work session.
