What should an AI agent keep from its past?

An AI agent working through a lengthy assignment accumulates more than conversation. Its history can contain search results, file locations, browser observations, failed attempts and intermediate reports. Feeding all of that back into a language model consumes tokens. Shortening it carelessly can remove something the next operation needs.

In a preprint submitted to arXiv on November 6, 2026, Qijia Zhuang, Rui Liu and Yuxiang Ren propose WAAC, a framework that compresses this history according to workflow dependencies. Rather than asking only which passages seem important, it asks which individual pieces of information support the work still outstanding, and how precisely they must remain visible.

Preserve the dependency, not necessarily the document

The researchers represent an agent’s executed operations as a directed acyclic graph: a network whose connections follow dependencies without looping backward. Nodes contain operations and their input and output fields. Connections record, for example, that a file path produced by one tool becomes an argument to another. Repeated attempts and updated artifacts remain separate events, preserving their order and versions.

This makes it possible to treat different fields in the same tool response differently. A path or status value might need to remain exact, while a long report could be represented by a recoverable reference.

WAAC uses four policies. EXACT exposes the original value. SUMMARY supplies a bounded account of specified semantic facts. POINTER supplies a reference to an externally stored object. DROP removes a field from the model-visible history. Crucially, these changes do not delete the original events or artifacts from runtime storage.

The policies are not interchangeable. A summary supplies readable meaning; a pointer supports recovery. If both capabilities are required and neither reduced representation suffices, the system retains the exact field instead.

Combining current needs with previous successes

WAAC builds a library from development trajectories that passed task-level verification. It replaces instance-specific values with typed placeholders and records recurring workflow structures alongside the field-retention policies used in successful runs. This library remains read-only during evaluation.

During execution, matching historical fragments provide initial policy proposals and potentially reusable dependency-analysis results. The online stage checks their applicability and recomputes requirements where dependencies have changed or are uncovered. It traces backward from outstanding obligations, including pending tool inputs, explicit plans, final-answer requirements and verification checks.

Historical policies can demand additional retention, but cannot weaken a requirement established by the current workflow. This distinction matters: previous experience does not directly permit more aggressive compression at an otherwise identical state. Its proposed benefit includes preserving useful context that helps later decisions and avoids repeated exploration.

Registered schemas and deterministic rules assign requirements and combine policies. Auxiliary models assist with structure extraction and summaries. Unresolved dependencies trigger exact retention; pointers require verified recovery evidence. These safeguards apply to the dependencies the system represents, not every possible future use of the history.

What the evaluations found

The study evaluated 147 PinchBench tasks and 200 Odysseys tasks. PinchBench covers tool-using activities such as research, programming and analysis. Odysseys tests long, multi-site workflows on the live web. The main WAAC agent used gpt-5.4-mini; some comparison methods required different checkpoints or inference adaptations.

On PinchBench, WAAC’s reported overall score was 87.6, compared with 84.3 for LLMLingua-2 and 72.6 for the uncompressed reference. Its input consumption was 7.174 million tokens versus 9.496 million for that reference, a reported reduction of 24.6%. Runtime was also lower than the uncompressed reference, although Keep-Last-N was faster than WAAC in the table.

On Odysseys, WAAC achieved perfect completion on 19.0% of tasks, compared with 15.0% for AgentFold. Average execution length was 58 steps versus 62 for AgentFold and 75 for the uncompressed reference. Even the best reported perfect-completion result therefore left most tasks short of full success.

Removing either historical matching or online dependency analysis reduced reported performance. These are end-to-end comparisons: changing retained context can change subsequent actions, so total token savings are not simply measurements of how much shorter one history becomes.

A bounded preservation claim

The results support workflow-aware retention in these evaluations, not a guarantee of safe compression across arbitrary tools or unseen workflows. The mechanism depends on registered schemas, identified dependencies and verifiable artifact recovery. If necessary information cannot fit the context budget, the design records that the budget is unsatisfiable rather than claiming every requirement can be preserved.

This is preprint research, not an established result. This account uses the supplied paper-text extraction and landing metadata; images were absent, and extraction can damage equations, tables and layout. The central proposal is nevertheless clear: compress an agent’s history according to how its remaining work uses that information, rather than treating all old text alike.