Hugging Face Open-Sources Funes: A Persistent Memory Layer for Coding Agents
Hugging Face engineer @ariG23498 released and detailed the open-source project Funes on September 7, providing a local, persistent memory layer for coding agents and directly addressing the well-known pain point of agents 'losing memory' when switching machines or sessions. The project quickly gained community attention, with @MaziyarPanahi, @dhruv2038 and @abhijithneil amplifying it.
Confirmed
- Funes is developed and open-sourced by the Hugging Face engineering team, with @ariG23498 as the original publisher
- Core positioning: a one-stop memory solution for agents, storing decision and execution traces and retrieving them on demand
- Supports no compaction at all, cross-machine persistent memory, and switching agents mid-session
- Compatible with mainstream coding agent tools including Claude Code, Codex, pi and Hermes, and can build memory from existing session logs
- @MaziyarPanahi noted that git cannot solve this: ignored files and system-level skills live outside the repository, requiring a dedicated memory-sync solution
Why it matters
- Coding agents today typically start from scratch in every new session, losing all context across machines and multi-agent collaboration—one of the biggest pain points in practice
- Funes turns session traces into retrievable persistent assets, stored locally without vendor lock-in, offering direct practical value for multi-machine workflows and long-term project maintenance
2026-09-07 ~ 2026-09-08 · 8 related posts
Primary sources
- [source] Hugging Face open-sources Funes, a durable memory layer for coding agents — ariG23498 · 2026-09-07
- [source] Hugging Face details Funes: turning agent session traces into durable, searchable memory — ariG23498 · 2026-09-07
- [source] Git can't sync agent memory: Hugging Face's funes tackles cross-machine agent state — MaziyarPanahi · 2026-09-07
5 near-duplicate retellings: ariG23498 · ariG23498 · MaziyarPanahi · dhruv2038 · abhijithneil