Task Advancement Memory, or TAM, organises GUI-agent memory around verified progress. The authors address long workflows in which useful steps span applications and successful local actions are mixed with failed attempts.
Record what changed in the task
During execution, Task Advancement State represents how verified progress changes the remaining goals. This is intended to let an agent build on work already completed instead of treating its history as an undifferentiated transcript.
After a task, advancement supervision identifies interactions that produced a verified local result and abstracts them into reusable operations. Those operations retain required inputs, expected outcomes and verification conditions. The method can preserve a useful local operation even when the overall task ultimately fails.
Reported improvements have different baselines
On MobileWorld GUI-only tasks, the state component alone raises success from 12.0% to 28.2%. The full system reports a 17.6% relative improvement over the strongest evaluated baseline and a 38.7% reduction in lifecycle tokens per solved task.
The relative gain and the state-only comparison are different results. They should not be merged into one improvement figure. Lifecycle tokens per solved task also concern the complete evaluated process, rather than just the tokens in a single action.
The paper additionally reports strong results on SPA-Bench in its single-app English Level 3 setting. That named setting is part of the finding, rather than a claim of best performance on every GUI task.
Verification gives memory its meaning
The framework’s practical idea is to distinguish an attempted interaction from an operation that achieved a checked result. An evaluator considering a similar memory design should examine how those verification conditions are defined and whether they still hold when an operation is reused.
The reported benchmarks motivate that comparison. They do not establish that storing more operations guarantees improvement on an unfamiliar application or that every earlier result can be transferred safely without checking its inputs.
Comments