Skip to main content
Back
40 XP

Lesson 3 · 12 min

Conversation history management

Three strategies for keeping conversation history within budget: sliding window, progressive summarization, and entity extraction.

The unbounded history problem

Every conversational AI has the same problem: conversation history grows without bound. Turn 1 is small. Turn 20 is large. Turn 100 exceeds the context window.

Naive solutions fail:

  • Keep everything → you hit the context limit and the oldest turns get truncated arbitrarily
  • Drop everything past N turns → you lose important earlier context abruptly, confusing the model

The right approach depends on what the conversation contains.