Memory compaction is the consolidation of many small memory entries into a smaller, current set, removing duplicates and superseded items.
Frequent checkpoints are good for safety and bad for tidiness. After a week of work, a project's memory may hold dozens of entries, some repeating each other and some overtaken by later decisions. Compaction merges them. Duplicate facts become one fact. A decision that was reversed is replaced by the reversal. Completed todos are closed. The result is a shorter memory that still says everything true about the project today.
Memory compaction is not the same as context compaction. Context compaction is done by the chat client to the live session and is lossy. Memory compaction is done to the durable store, deliberately, and should keep the underlying entries recoverable. ChatGPT's June 2026 move to an auto-rewritten memory summary is a form of compaction the user cannot control or inspect. Compaction that runs on files you can open is auditable; compaction inside a vendor's store is not.
How Central Brain handles memory compaction
memory_compact tool replaces the project brief with an updated summary and marks older notes as compacted rather than deleting them. The notes stay as local files you can read.