Grow-and-Refine Context Maintenance

Keep an incrementally growing bullet context compact by appending new bullets, updating existing ones in place, and periodically de-duplicating via semantic embeddings, either proactively or lazily.

Incremental delta updates only add; grow-and-refine is the complementary housekeeping step that keeps an itemized bullet context relevant as it grows.

Mechanics:

  1. Bullets carrying new identifiers are appended.
  2. Existing bullets are updated in place — e.g. their helpful/harmful counters incremented.
  3. A de-duplication pass compares bullets by semantic embedding similarity and prunes redundant ones.

Refinement timing is a latency/accuracy knob:

  • Proactive — refine after every delta; the context is always compact, at the cost of extra work per step.
  • Lazy — refine only when the context would exceed the window; cheaper on average, but the context carries redundancy until the trigger fires.

The combination aims for contexts that expand adaptively, stay interpretable (each bullet is human-readable and traceable to its usage counts), and avoid the variance of monolithic rewriting. Embedding-based de-dup is a non-LLM step, so it cannot introduce the summarization loss that context collapse describes — though a too-loose similarity threshold can still merge bullets that differ in a detail that matters. Compare the eviction-driven alternative, where pressure on the window triggers summarization (memory pressure eviction), which trades this preservation for a hard size bound.

Knobs and sensitivity (App. A.6; FiNER only, base 70.7):

  • De-dup similarity threshold (how aggressively a new bullet merges into an existing one): 50% → 77.0, 70% → 73.9, 90% → 78.6. Every setting beat base, and the authors call the effect mild. But 70% was about 4.7 points below 90%, so sweep this knob on your own task rather than assume any value works.
  • Pruning trigger (maximum context length before older or low-utility bullets are merged or pruned): 10K → 78.6, 50K → 78.4, 100K → 78.3. Stable on this task. The authors' interpretation (not separately measured) is that pruning mostly removes stale or harmful fragments and keeps the core reusable strategies.

The refine step can also filter entries whose metadata flags them as potentially harmful. The authors call it a "first line of defense" in context adaptation noise robustness.

Sources

  • ACE (Zhang et al.), §3.2, App. A.4, A.6