A memory checkpoint is a deliberate save of an AI session's current state (decisions, facts, open questions, todos) to durable storage.
The idea comes from computing generally: a checkpoint is a snapshot you can return to. In AI memory, a checkpoint is written by the agent or the user at a natural pause, such as after a decision is made, a bug is found, or a task is handed off. It records what changed, not the whole transcript. A good checkpoint is short, structured and dated, so the next session can read it in seconds rather than re-reading hours of chat.
Checkpoints solve two problems. First, context compaction: a summary written by the client may drop details, but a checkpoint written to disk does not. Second, tool switching: a checkpoint saved from Claude can be read by Copilot or Codex if the store is shared. The discipline is to checkpoint often and keep each entry small. Consolidation later (memory compaction) merges the small entries into a clean current view.
How Central Brain handles memory checkpoint
memory_checkpoint tool appends one note at a time (a decision, fact, question, todo or summary) as you go, and memory_recall finds them again when a new session starts.