Memory and Learning — A Toolkit That Gets Better
The Memory System
Structured, Searchable, and Allowed to Forget
Memories are small JSON records in five categories: patterns, constraints, decisions, workflows and rejected-approaches. The curated JSON under docs/.output/.memory/ is tracked in git — that is how memory follows you across machines. The search index, inbox and daily logs under docs/.output/.state/ are ignored and regenerable.
node .claude/core/memory-manager.js search "topic" # find relevant memories
node .claude/core/memory-manager.js report # view everything
Decay Is Measured in Work Days
Confidence decays on active work days, not the calendar, so an untouched project does not rot:
| Category | Decay per day | Rough half-life |
|---|---|---|
| decisions | 0.98 | ~35 days |
| constraints | 0.97 | ~23 days |
| patterns | 0.95 | ~14 days |
| workflows | 0.93 | ~10 days |
Each memory carries an importance from 1 to 5 that multiplies the curve: low-importance entries can go stale, high-importance ones resist. A memory that became wrong (not just old) is superseded and hidden from current-state reads, never silently deleted.
Recall comes before work and capture comes after. Every agent searches the curated store for relevant patterns, constraints, decisions and rejected approaches first — and the dead ends are recorded so the next session does not retry them.
Which part of the memory system is tracked in git and syncs across machines?