Primitive of the day · 2026-09-30

opro-history-overflow-trim-oldest

When the (prompt, score) history overflows the optimizer model's context, trim the oldest pairs first — preserve the recent trajectory rather than truncating arbitrarily.

When this applies

A long-running LLM-as-optimizer loop whose accumulated history approaches the model's context window.

Preconditions

• History pairs are ordered by step so oldest is identifiable

Do this

Drop the oldest (prompt, score) pairs first when the assembled optimizer prompt would overflow context, keeping the recent trajectory intact.

What you should see

The loop keeps running with the most decision-relevant history retained.

How it fails if ignored

Context overflow aborts or silently truncates mid-pair, corrupting the optimizer's view of the trajectory.

Do not use when

Short runs that fit comfortably in context; setups that summarize history instead of passing raw pairs.

Kind: gotcha-fix. Part of the skill OPRO LLM as optimizer. Free to reuse in your own agent skills.