Hugging Face Launches Funes: Persistent Cross-Agent Memory for Claude Code and Codex
Hugging Face released funes, an open-source CLI tool providing durable, cross-agent working memory across Claude Code, Codex, pi, and Hermes. It deterministically indexes local session traces into Lance datasets, reducing context reloading overhead up to 8x compared to manual handoffs.
Impact: Medium
Why it matters
Install funes to let coding agents query past execution decisions and debugging context across sessions without repeatedly burning tokens on context compaction.
TL;DR
- 01Funes operates locally as a single binary with zero external machine learning dependencies, using Lance datasets for persistence.
- 02Cross-agent compatibility allows traces generated in Claude Code to be queried and leveraged immediately within Codex or Hermes.
- 03Retrieval-based memory proved 8x cheaper than written handoffs in comparative evaluations, avoiding failure modes of context compaction.
Key facts
- Cost reduction vs written handoff
- Up to 8x cheaper in evaluated tasks
- Supported coding agents
- Claude Code, Codex, pi, Hermes
- Underlying storage engine
- Local Lance dataset
Solving Context Loss Across Developer Workstations
When a terminal agent session finishes, architectural context and troubleshooting rationales are lost. Funes addresses this by transforming passive execution logs into an active memory layer for Claude Code, Codex, pi, and Hermes. Instead of summarizing text at write time, funes stores raw turns and exact evidence, allowing agents to trace every retrieval hit back to the originating session and timestamp.
Local Hybrid Retrieval Pipeline
The architecture relies entirely on local execution without hosted model dependencies:
- Incremental indexing: Turns are parsed and embedded locally as they complete, backfilling deep history in bounded steps.
- Search fusion: Combines local vector search and BM25 indexing in a Lance dataset, with cross-encoder candidate reranking.
- Cost reduction: Benchmark tests show recall is up to 8x cheaper than maintaining written handoffs and avoids compaction loss where summaries erase necessary technical nuance.
Installation and Multi-Agent Setup
Install the standalone binary and attach it directly to an existing agent:
curl -fsSL https://huggingface.co/buckets/huggingface/funes/resolve/install.sh | sh
funes add claudeOnce added, agents access recall and get commands internally during regular workflows. To query memories interactively without agent modification, run funes ask claude "query".
Try it in 2 minutes
curl -fsSL https://huggingface.co/buckets/huggingface/funes/resolve/install.sh | sh
funes add claude # or: codex, pi, hermesbash
✓ When to use
- When switching across multiple machines or toggling between Claude Code and Codex on the same project.
- To prevent agent context windows from bloating during multi-day debugging investigations.
✕ When NOT to use
- When running strictly ephemeral, throwaway coding tasks where retaining rationale across sessions has zero value.
- In environments subject to strict offline security policies forbidding new binary installation scripts.
What to do today
- Install funes on local workstations to enable agent trace indexing across Claude Code and Codex.
- Execute funes add on active coding CLI tools to expose recall and get commands inside the loop.
- Optionally bind local memory to a private Hugging Face dataset to sync agent context across development machines.
Sources