# Hugging Face introduces Funes for portable coding-agent memory > Funes indexes coding-agent histories locally and can share them through a Hugging Face dataset, preserving source turns behind retrieved decisions. Canonical URL: https://www.devobs.io/news/news-hugging-face-funes-agent-memory/ By: Lucas Vale Published: 2026-09-06T11:58:54.633Z Updated: 2026-09-06T11:58:54.633Z Event date: 2026-09-03 Section: AI Hugging Face introduced Funes on September 3, an open-source tool for turning coding-agent session histories into searchable memory. The [announcement](https://huggingface.co/blog/funes) describes local indexing and retrieval, with optional sharing through a Hugging Face dataset controlled by the user. ## Retrieval preserves the original session Funes embeds and reranks locally, returning source text with references to its originating session and turn. It combines vector and keyword search over a common representation of supported agent traces. Shared datasets are private by default, and the announcement describes redaction during indexing plus another scan before publishing. This design makes old reasoning available to a later session without requiring the developer to manually reconstruct every past investigation. ## Keep remembered reasoning tied to current reality Our analysis: retrieved history should be treated as evidence about an earlier decision, not as an instruction that automatically overrides the present task. A once-correct conclusion may become stale after a dependency change, a migration or a revised product requirement. Useful memory evaluation should include questions whose answers changed over time. Check whether the system retrieves the later correction, preserves the earlier rationale and makes the chronology understandable. Finding a highly relevant old passage is not enough if it points the agent toward a superseded solution. Sharing also expands the audience for session content. Review which projects and conversations enter a shared dataset, and define retention and deletion expectations. Automated secret scanning is a helpful control, but it cannot establish that every passage is suitable for every teammate. The release offers a concrete way to make agent work inspectable across sessions. Its value depends on retrieving the right context and then checking that context against the current repository and the user’s present authorization. ## Source references -