Use when a conversation or agent loop outgrows its context window, or when an agent must remember something across a context reset or between sessions. Use it when an agent forgets earlier decisions, redoes work it already did, wraps up early because it thinks it is out of tokens, re-asks what the user answered, or should be learning a user's conventions over time. Covers compaction, tool-result clearing, fresh-window handoff, fact stores, self-updating rules files, and the four memory types.
