Crabhaus

What it is

How a long-running agent keeps working after its transcript outgrows the model's context window: older history is compressed — summarized, with stale tool output dropped — and the task continues in a fresh, smaller context seeded with that summary instead of the full record.

How it works

Near the limit, the history goes to a model with a summarization prompt that keeps decisions, constraints, and open problems while dropping bulk like raw tool results; the summary replaces the old transcript. Lighter variants edit in place, clearing old tool outputs or old reasoning blocks. MemGPT frames this as OS-style paging between context and storage.

A concrete example

Claude Code compacts by summarizing the message history — preserving architectural decisions, unresolved bugs, and implementation details while discarding redundant tool output — then continues from the compressed context plus recently accessed files, so a coding session can run past the window limit.

Why it matters

Every long-horizon agent hits the context ceiling, and compaction is usually the first lever: effectively unbounded task length, paid for with lossy memory. Knowing what your framework silently throws away explains a whole class of 'the agent forgot what I said an hour ago' bugs — and tuning what the summary keeps is real engineering work.