arxiv:2609.34565

FlowState: Execution State as Memory for Long-Horizon LLM Agents

Published on Oct 8

· Submitted by

bangyan li on Oct 21

ant-international

Upvote

Authors:

Bangyan Li ,

Abstract

FlowState improves long-horizon agent performance by maintaining structured, revisitable execution memory with incremental updates and progressive access, reducing token use while increasing success rates.

Generated by thinkingmachines/Inkling-Small

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. To address these challenges, we propose FlowState, which treats execution state as memory that can be retained and revisited across requests, unifying current decision-making with the reuse of historical information. FlowState preserves semantically typed state nodes, their relations, and references to raw tool observations, separating persistent retention from on-demand access. Within a single execution loop, Incremental State Update (ISU) maintains the current state based on new inputs and feedback, while Progressive State Access (PSA) progressively reveals historical states and supporting evidence as needed during reasoning. Together, these mechanisms enable agents to reassess prior decisions in light of new information and guide subsequent actions. Compared with a full-context baseline using the same DeepSeek-V4-Flash model, FlowState improves the average success rate on MemoryArena and the average pass rate on τ^3-Bench by 4.55 and 13.95 percentage points, respectively, while reducing total token consumption by 43.2% and 40.6%. These results demonstrate the performance and efficiency advantages of FlowState on long-horizon tasks.

View arXiv page View PDF Add to collection

Community

li-fly

Paper author Paper submitter 4 days ago

We explore how long-horizon LLM agents can preserve useful information without repeatedly loading their entire interaction history. FlowState maintains structured execution states linked to original tool observations, updates them as tasks progress, and retrieves relevant states and evidence on demand.
Using the same DeepSeek-V4-Flash backbone, FlowState improves average success rate on MemoryArena by 4.55 percentage points and average pass rate on τ³-Bench by 13.95 percentage points over a full-context baseline, while reducing total token consumption by 43.2% and 40.6%, respectively. We welcome questions and feedback, especially on agent memory and long-horizon task execution!

O96a

4 days ago

I've been running the cheap version of this for months — a typed JSON state file the agent reads and writes between requests instead of replaying the transcript — so the retention half is already settled for me, and it costs nothing. What I want numbers on is invalidation. Typed nodes go stale the moment the environment moves: a file gets edited, an API changes shape, a tool starts returning a different schema, and the agent keeps reasoning off a node that was true forty steps ago. Stale state fails quieter than a long context does, because nothing in the prompt looks wrong. So inject a mid-run environment mutation and report whether the agent re-verifies the affected nodes or just trusts them, and whether the state graph carries a timestamp or provenance field the model can actually condition on. If the answer is "the model decides when to revisit," that's the same vibes-based memory I have now, with a schema and extra steps.

li-fly

Paper author 4 days ago

Thanks for the concrete suggestion. We agree that state validity in dynamic environments deserves a dedicated evaluation. FlowState aims to let agents not only retain state, but also revisit earlier judgments and their supporting evidence as a task unfolds, and record revisions through incremental updates. Compared with simply persisting structured state, this provides more explicit support for information reuse and revising judgments, with gains in task success and efficiency in our evaluations.
We currently retain source-request information and evidence references, but do not enforce verification timestamps or automatically propagate invalidation; when to re-verify still depends on the model. In future work, we hope to test scenarios where the environment changes during execution, evaluate how reliably agents re-verify affected states and revise dependent judgments, and explore more robust invalidation mechanisms.