Language agents must fit accumulated evidence into a limited token budget. Keeping raw records preserves exact details but can leave relevant evidence out; consolidation compresses records to cover more but can lose query-critical details. This work formalizes the choice by decomposing each consolidation operator's utility into a coverage effect on evidence retention omits and a signed replacement effect on raw evidence that already fits.
Using the LongMemEval and LoCoMo benchmarks, the authors found a consistent budget-dependent crossover. On LongMemEval, consolidation improved accuracy by up to 48 percentage points under tight budgets, while retention was preferable under loose budgets. LoCoMo reproduced the crossover at a smaller absolute budget, matching its shorter evidence. Across both benchmarks, cross-note Abstract and Merge generally outperformed local Rewrite when compression was needed.
The result matters because it replaces a fixed token threshold with a relative budget pressure rule, and it is operationalized through OAS, a lightweight learner that estimates action utilities from pre-generation features and uses held-out harm calibration to avoid harmful replacements.