Before the recent upgrade to Claude Code's memory system, I shared a plan for adding memory to Claude Code. Recently, a new round of upgrades was implemented, with the core change being the separation of "what to remember, when to remember, and where to store" into three layers:

**Capture Layer** — Defines which signals are worth recording: over 15 minutes of troubleshooting, counterintuitive technical discoveries, trade-offs in strategic decisions. Not all conversations are worth remembering; key is filtering.
**Timing Layer** — Controls the writing rhythm: after completing each task, record progress; after each commit, update project context; upon detecting leave signals ("That's it," "Leaving"), immediately save everything. No manual trigger needed.
**Routing Layer** — Ensures data is written to the correct location: technical pitfalls → project; reusable experience → daily progress. Uses ChromaDB for vector storage + BM25 keyword search as a plain text backup. Implements double writing (database + file); indexes are incrementally synchronized rather than rebuilt entirely.

Each layer manages one aspect: adding new capture rules doesn't require changing the writing logic; changing storage locations doesn't affect signal recognition. Claude Code's hooks mechanism is naturally suited for this event-driven pipeline.

**Result:** Cross-session memory is finally stable. Claude can remember last week's troubleshooting experiences, and during debugging, it will review historical records before acting.