Claude Code auto memory gets Dream pass
Claude Code stores per-project learnings in ~/.claude/projects/<project>/memory/ and only loads the first 200 lines of MEMORY.md at session start. A GitHub system prompt suggests a hidden Dream consolidation pass that reorganizes those notes into topic files, updates the index, and prunes stale memory.
This is the right kind of unglamorous AI feature: memory only works if it self-curates, and Claude Code is treating context like a maintained file system instead of a chat log. The upside is better continuity across sessions; the risk is that hidden rewriting can turn into hidden forgetting if the index gets sloppy.
- –`MEMORY.md` is an index, not the warehouse, so detailed notes have to live in separate topic files to stay under the 200-line startup ceiling.
- –The Dream pass appears to merge fresh signal, delete contradicted facts, and demote verbose entries, which is smarter than just appending more notes.
- –Users are already bumping into the line limit, so memory quality now depends as much on pruning discipline as on retrieval.
- –The editable `/memory` workflow is a big trust win because users can inspect and fix what Claude remembers instead of hoping the model got it right.
DISCOVERED
66d ago
2026-03-24
PUBLISHED
66d ago
2026-03-24
RELEVANCE
AUTHOR
DIY Smart Code
