Finished reasoning keeps its seat
A language model agent working over a long horizon accumulates a record of its own reasoning as it goes. How it judged, why it took that branch. That record stays in the context even after the decision has been executed and its outcome observed. As the context grows, inference costs rise, and the longer the task the more that tells. The agent keeps paying, on every step, for reasoning that was settled long ago.
The obvious response is to delete the old reasoning. The authors argue this is not a summarisation problem. Unlike compressing a finished chain of thought, removing an agent's history changes what it does next. The deletion sends it down a different branch, to different observations, to different judgements. So the question is not how much can be cut, but when reasoning can be safely forgotten. The primary source for this article is a preprint that has not been peer reviewed.
Keep the actions, drop the reasoning
The proposal narrows to one idea. Preserve actions, tool calls and observations exactly, and rank only the blocks of reasoning for removal. The authors call this interaction-aware compression[1]. It requires no training and decides online, while the task is still running.
The ranking uses an uncertainty measure produced by a frozen proxy model. Which is to say, the question of which reasoning block earns its place is answered by a fixed yardstick rather than by a separately trained judge. That the method is light is its advantage and equally its limit. A fixed yardstick cuts by the same criterion no matter how the character of the task changes.
Static compression
Shorten a chain of thought after it is complete. Compare the original against the compressed version and you can measure what was lost. Nothing you delete produces a new action.
Interaction-aware compression
Delete history while the task is still running. The deletion changes the next action, and the changed action produces a different history. There is no counterpart left against which to measure the loss.
What the abstract states
The authors evaluate on 260 WorkBuddyBench tasks. Average reward rises from 0.699 to 0.718, while input, output and cache-read tokens fall by 25.5%, 14.4% and 33.3% respectively. The claim is that cost went down without the score going down with it.
The abstract also records a more interesting observation. Deleting reasoning locally changes total computation non-linearly, an effect the authors name trajectory amplification. The saving is not proportional to what was removed, because the subsequent interaction changes, which makes both the reduction and any increase hard to predict. Further analysis, including representation probing and activation patching, suggests that historical reasoning becomes more replaceable once the task-relevant derived state has been externalised into code, files, tool outputs or feedback from the environment. That is the extent of the abstract, and it is the authors' claim.
What this looks like in practice
Applied to work that passes through several hands, the argument reads like this.
Translated into a handover
Successor: "Why was this condition set this way?"
Predecessor: "There are no minutes, but the condition itself is written in the specification."
Successor: "It behaves as specified, so I do not need the history."
Where the conclusion has been written down outside, the reasoning that led to it can be discarded. Where it has not, the reasoning is the only record. That distinction is what the paper is proposing, and it places the condition for forgetting on whether the state was externalised. Read the other way: in work with weak means of writing things down, the volume of reasoning that must not be cut grows.
What is not new, and what cannot be read
Discarding working memory that is no longer needed is not a new idea. Optimisations that throw away unwanted intermediate results are almost as old as computing. What this paper contributes is not the idea but the willingness to face a setting where deletion alters subsequent behaviour, and the choice to base the deletion decision on a fixed, training-free yardstick.
Much cannot be read from the abstract. What share of reasoning was removed. How the score moved across different categories of task. How many cases showed trajectory amplification working against the agent. None of these conditions are stated. The concentration of the evaluation on a single benchmark deserves particular care: change the character of the work and the amount of derived state that gets externalised changes with it, and so should the share of reasoning that is safe to drop.
Why this subject is clustering now
A paper published in the same window presents a framework for driving epidemic models with artificial intelligence agents[2]. The subject matter is remote, but the question overlaps. When an agent is handed a long process, what does it hold inside itself and what does it delegate to machinery outside? On the epidemic side, the computation lives in an external model and the agent handles invocation and interpretation. On the forgetting side, it is precisely because state has been externalised that reasoning can be dropped. Both are faces of the same question about where the boundary between an agent's inside and outside should be drawn.
For this subject the only channel standing is the cluster of papers asking the same question. No reader votes, no implementation following, no press pickup appear in the collected material. This site measures the spread of a subject through independent channels, but a cluster shows only that researchers are converging on a question, not that an answer has arrived. Like vote counts, it sits on a different axis from correctness.
Where this connects to material review
Here is a point that collides directly with regulated practice. The reasoning history this paper treats as a cost is, in material review, the record itself. Why this wording was accepted, why that figure was sent back. Without the process of judgement retained, there is no way to confirm afterwards that the same standard was applied. An audit asks not for the conclusion but for the path to it.
Invert the paper's claim, though, and something usable for review design appears. The condition that forgetting is safe once derived state has been externalised is, turned around, a demand to write the grounds of a judgement down outside your head. Rather than keeping only the result of a comparison made mentally, put into the ledger which passage of which document was checked against what. Written to that depth, a judgement can be reconstructed without preserving the reasoning verbatim.
At the same time, if a design that deletes history to lower cost is carried into a review tool, the fixed yardstick becomes the risk. If no record survives of which reasoning was discarded, the deletion itself falls outside the scope of audit. A system that does not record what it threw away cannot explain the judgements it threw away with it. Cost is measurable; the price of vanished grounds is measurable only after the fact.