Apple Researchers Propose SELFCOMPACT for Adaptive Context Management in LLM Agents

Long agent traces composed of chains of thought and tool calls accumulate stale content that eventually outgrows the context window. Existing scaffolds mitigate this with fixed-interval compaction triggered at a token threshold, but such triggers ignore trajectory structure, risking the discard of partial results mid-derivation or mid-search.

Adaptive Compaction via Self-Decision

Researchers propose SELFCOMPACT, a scaffold that allows the model itself to decide when and how to compact context. The approach pairs two inference-time elements: a compaction tool the model invokes to summarize accumulated context, and a lightweight rubric specifying when to fire (e.g., a sub-task has resolved) and when to suppress (e.g., mid-derivation). Both components are necessary; the tool alone is unevenly used across open-weight models, while the rubric alone cannot act.

Performance and Cost Efficiency

Empirical results on six benchmarks (competitive math and agentic search) and seven models show that SELFCOMPACT matches or exceeds fixed-interval summarization at a fraction of the token cost. The method improves over a no-summarization baseline by up to 18.1 points on math and 5–9 points on agentic search, while achieving 30–70% lower per-question cost. The study highlights a meta-cognitive gap: unprompted models cannot reliably detect when their context is degrading, but a lightweight rubric closes this gap without requiring fine-tuning or external supervision.