Each turn injects the memory directory tree — paths, one-line summaries, updated-at times — rather than full file bodies or embedding matches; the model pulls details with memory search or direct reads. This tells the Bot what it remembers without spending its context budget, and keeps a semantic-retrieval stack out of the PoC.
Consequences
The tree is capped (1000 paths) and folders with many files are folded into counts; semantic search is revisited only if keyword search proves insufficient (PRD §5.5 rule M3, appendix D).