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

FlowState. Execution State as Memory for Long-Horizon LLM Agents. By Minghao Li and seven others, from Ant International, Ant Group. Published on the eighth of October. No code repository is linked to it.

A long-horizon task keeps asking an LLM agent to reach back into earlier interactions. The two obvious answers both cost something. Keep the full history, and context costs climb. Compress it, and you risk throwing away a detail that will matter later. The deeper difficulty is timing: whether a piece of history is relevant often only becomes apparent as the task unfolds, so there is no good moment to decide in advance what to keep.

FlowState reframes the question. Rather than treating memory as a transcript to be kept or shrunk, it treats the agent's execution state as the memory, a state that can be retained and revisited across requests. Deciding what to do now and reusing what happened earlier become one activity rather than two. The key separation is between persistent retention and on-demand access: things are kept, but only opened when needed.

What is kept are semantically typed state nodes, the relations between them, and references back to the raw tool observations. Inside a single execution loop, two mechanisms work on that structure. Incremental State Update, or ISU, maintains the current state from new inputs and feedback. Progressive State Access, or PSA, reveals historical states and their supporting evidence progressively, as reasoning calls for them. Between the two, an agent can reassess earlier decisions when new information arrives, and let that reassessment steer what it does next.

The comparison holds the model fixed: a full-context baseline and FlowState both run on DeepSeek-V4-Flash. On MemoryArena, FlowState raises the average success rate by four point five five percentage points. On tau three Bench, it raises the average pass rate by thirteen point nine five points. And it does both while spending less: total token consumption falls by forty-three point two percent on the first benchmark and forty point six percent on the second.

Against a baseline that simply keeps everything, both numbers move the right way at once: better results and around two fifths fewer tokens. That pairing is the substance of the claim. Keeping the history as structured state you can return to, instead of text you must carry, is what lets the agent remember more while reading less.