Sliding Window

Retain only the most recent N conversation turns in the agent's context. Oldest turns are dropped entirely. This caps the prompt size at a fixed bound regardless of session length.

Trade-offs

• Pros: cheap, deterministic, easy to implement.
• Cons: loses early decisions and long-range dependencies; fragile on long debugging or refactoring tasks where the original goal or constraints were stated at the start.

When to Use

Best as a fallback or combined with summarization. A pure sliding window is usually insufficient for coding agents working on multi-file tasks.

See Also

• Rolling Summarization — combine with sliding window to preserve state from dropped turns.
• Compaction — sliding window is one of several compaction strategies.

Tools

• LangChain ConversationBufferWindowMemory — keeps only the last K interactions in the buffer.