Overview
On September 3, 2026, Hugging Face released Funes, an open-source memory layer for coding agents, designed to address the issue of coding agents losing context in cross-session, multi-file, and multi-repository tasks. Funes consolidates the exploration and decisions of each session into retrievable long-term memory, providing coding agents with a standardized memory layer that enables sharing of codebase understanding across different agents and between sessions of the same agent. It defaults to local Lance dataset storage and optionally syncs with Hugging Face datasets for multi-machine sharing. Funes is compatible with multiple agent frameworks such as Claude Code, Codex, pi, and Hermes, without locking users into a single vendor. Its hybrid retrieval strategy combines vector retrieval, BM25 keyword search, cross-encoder reranking, and time decay. The official claim is that its cost is about 1/4 to 1/8 of handwritten handover solutions.
Key Features
- Open-source memory layer: As an open-source project, Funes provides a standardized long-term memory layer for coding agents, supporting context persistence across sessions, files, and repositories, allowing agents to accumulate and reuse exploration and decision-making experience.
- Local and cloud storage options: It defaults to local Lance dataset storage, ensuring data privacy and fast access; optionally syncs with Hugging Face datasets for multi-machine memory sharing, suitable for team collaboration or distributed workflows.
- Cross-framework compatibility: Compatible with multiple coding agent frameworks such as Claude Code, Codex, pi, and Hermes, without binding to a specific vendor. Users can flexibly switch between different tools while retaining a unified memory layer.
- Hybrid retrieval strategy: It combines vector retrieval with BM25 keyword search, reranks results using cross-encoders, and incorporates a time decay mechanism (recency) to ensure retrieval results are both relevant and timely.
- Significant cost efficiency: According to the official statement, compared to handwritten handover documentation, Funes saves tokens and time, with costs approximately 1/4 to 1/8 of handwritten solutions, greatly reducing maintenance overhead for coding agents.
Use Cases
- Cross-session codebase understanding: When an agent handles the same project across multiple sessions, Funes retains memory of the code structure, avoiding repeated exploration.
- Multi-file task collaboration: In complex tasks involving multiple files, agents can retrieve previous decisions on related files to maintain consistency.
- Multi-repository knowledge sharing: When switching between multiple repositories, agents can access historical context for each repository through Funes, improving cross-repository development efficiency.
- Team multi-machine collaboration: Through cloud synchronization, agent instances on different machines can share memory, suitable for team collaboration or remote development scenarios.
- Smooth framework migration: When users switch from Claude Code to tools like Codex, the memory layer retained by Funes can be seamlessly migrated, reducing relearning costs.
Pros
- Open-source and free to use, lowering adoption barriers, with community participation for improvement.
- Flexible storage, balancing privacy and sharing needs with local and cloud sync.
- Compatible with multiple mainstream agent frameworks, avoiding vendor lock-in.
- Hybrid retrieval strategy enhances the accuracy and timeliness of memory retrieval.
- Costs are significantly lower than handwritten handover solutions, saving tokens and time.
- Official blog provides detailed technical explanations, facilitating developer understanding and deployment.
Pricing
Funes is an open-source project and can be used for free. Specific pricing details (such as storage costs for cloud sync) are not specified in the official release; please refer to the official website.
Summary
Funes is an open-source coding agent memory layer launched by Hugging Face, achieving cross-session and cross-framework context persistence through local Lance storage and optional cloud sync. Its hybrid retrieval strategy combines vector and keyword search with time decay, ensuring efficient and useful memory. The official claim is that costs are only 1/4 to 1/8 of handwritten handover solutions, making it suitable for coding agent scenarios requiring long-term memory.
Version History
- Funes released (open source) (2026-09-03): Hugging Face released Funes, an open-source memory layer for coding agents: local Lance dataset storage with optional Hugging Face dataset sync, framework-agnostic (works with Claude Code, Codex, pi, Hermes and others), and hybrid retrieval (vector + BM25 + cross-encoder reranking + recency decay). Hugging Face claims it is 4-8x cheaper than handwritten handoff documents.