What they did

The authors define memory management for language agents as a budget-dependent operator selection problem. They decompose each consolidation operator's utility into a coverage effect (evidence omitted by retention) and a signed replacement effect (from replacing raw evidence that already fits), explaining why the preferred action changes with relative budget pressure. They implement OAS (Offline Abstraction-Safety), a lightweight learner that estimates action utilities from pre-generation features and uses held-out harm calibration. The system is evaluated on LongMemEval (long-term memory benchmark) and LoCoMo (long-context memory) under varying context budgets.

Key findings

• On LongMemEval, consolidation improves absolute accuracy by up to 48% under tight budgets, while retention is preferable under loose budgets.
• LoCoMo replicates the same crossover at a smaller budget, consistent with its shorter evidence.
• Cross-note abstraction and merging generally outperform local rewriting when compression is necessary.
• OAS selects the optimal operator per query, achieving the best of both retention and consolidation.

Why it matters

This provides a principled framework for memory management in language agents, moving beyond ad-hoc heuristics. It highlights that the optimal strategy depends on budget pressure, a crucial insight for designing scalable agent systems that must operate under context window limits and inference costs.

Caveats

The study is limited to specific benchmarks and three consolidation operators (Merge, Abstract, Rewrite). The OAS learner relies on held-out calibration data, which may not generalize to all domains or query types. Scalability to very large memories and dynamic, multi-turn interactions remains unexplored.