Context Adaptation

Improve an LLM system by iteratively revising its inputs — system prompt, memory, strategies, evidence — from natural-language feedback on execution, instead of updating model weights.

Context adaptation is the self-improving branch of context engineering. The context is not assembled once. It is a learned artifact that the system revises from its own experience. The standard loop is:

  1. An LM inspects the current context together with signals such as execution traces, reasoning steps, test or validation results.
  2. It writes natural-language feedback on how the context should change.
  3. The feedback is folded back into the context, and the cycle repeats.

Representative methods differ mainly in what the adapted context is. Reflexion adapts per-task reflections for agent planning. TextGrad applies gradient-like textual feedback to prompts. GEPA iteratively rewrites a prompt from execution traces. Dynamic Cheatsheet accumulates a test-time memory of strategies. ACE's generator–reflector–curator loop evolves an itemized playbook.

There are two operating settings:

  • Offline: optimize a system prompt on a training split, then ship it.
  • Online: adapt test-time memory continuously while the agent runs.

Why adapt context rather than weights:

  • Interpretable: developers can read, diff, and audit the learned artifact.
  • Fast to integrate: new knowledge takes effect at runtime, with no training run.
  • Portable: a context can be shared across models or modules of a compound AI system.
  • Long-context models and context-efficient inference (e.g. KV-cache reuse) make large adapted contexts increasingly practical to deploy.

Two characteristic failure modes come from how the context is revised: brevity bias (the optimizer converges on short, generic instructions) and context collapse (monolithic rewrites erase accumulated detail). Both argue for incremental delta updates over regenerating the context.

The authors frame context adaptation as a continual-learning mechanism for distribution shift and scarce training data. They present it as a flexible, generally cheaper alternative to fine-tuning; that is a framing, not a measured comparison in the paper. Because the learned artifact is readable, it also supports selective unlearning of specific items (see itemized bullet context).

Sources

  • ACE (Zhang et al.), §1–2.1, §5