Context compaction is the process of summarizing or trimming older parts of a conversation so a model can keep working within its context window.
When a long session approaches the context limit, the client (Claude Code, Codex, ChatGPT, Copilot) compresses earlier turns into a shorter summary and discards the originals. The model then continues from the summary. This keeps the session alive but is lossy by design. Exact wording, tool outputs, file paths and small decisions are the first things to go, because a summary keeps what looks important at the moment of compaction, not what you will need later.
Compaction is a client feature, and each client does it differently. Some announce it; some do it silently. What they share is that the summary lives only inside that session. If you open a new session, the summary is gone too. The practical defense is to write important state to durable memory before compaction happens, then reload it afterwards. A compacted session with a good external checkpoint loses almost nothing. One without it loses whatever the summarizer skipped.
How Central Brain handles context compaction
memory_checkpoint live in the project's memory files, not the session, so they survive compaction. Your session got compacted. Your memory didn't.