FlowState keeps execution state as memory for long-horizon agents
Original title: FlowState: Execution State as Memory for Long-Horizon LLM Agents
Against a full-context DeepSeek-V4-Flash baseline, authors report higher pass rates and about 40% fewer tokens.
The paper proposes FlowState, which stores execution state as reusable memory for long-horizon LLM agents instead of keeping or compressing the full history. Incremental State Update maintains the current state, while Progressive State Access reveals earlier states and tool evidence during reasoning. Using the same DeepSeek-V4-Flash model, the authors report gains on MemoryArena and τ^3-Bench with lower token use.
Unconfirmed: Improvements are the authors’ comparisons against one full-context baseline.
- State nodes are semantically typed and can reference raw tool observations.
- MemoryArena average success rises by a reported 4.55 percentage points.
- τ^3-Bench average pass rate rises by a reported 13.95 percentage points.
- Reported token cuts are 43.2% on MemoryArena and 40.6% on τ^3-Bench.
- AI models
- DeepSeek-V4-FlashFlowState
From the original article
1 min read at huggingface.co
Long-horizon tasks require LLM agents to continually draw on information from earlier interactions. However, retaining the full history increases context costs, while compressing it risks losing details needed later, and the relevance of historical information often becomes apparent as the task progresses.
By Minghao Li, Bangyan Li, Zifan Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.