| name | sediment |
| description | Capture reusable knowledge after milestones. Systematically review what was learned and persist it to the right layer. Keywords: sediment, crystallize, capture, persist, remember, learned, milestone, retrospective. |
Sediment
Persist valuable, reusable knowledge to the right layer after completing a task or milestone.
When to Use
- After fixing a non-trivial bug (root cause pattern worth remembering)
- After a design decision that affects future work
- After discovering an undocumented behavior in a dependency
- After establishing a new workflow or convention
- When the user explicitly asks to sediment/capture/remember
Workflow
1. Identify what's worth keeping
Ask: is this stable, reusable, and verified? If it's speculative or session-specific, skip it.
2. Check for stale entries
Before writing new entries, scan existing memory, skills, rules, and CLAUDE.md for anything this session invalidated or superseded. Update or delete stale entries first.
3. Route to the right layer
| What | Where | Example |
|---|
| Project-specific facts, known bugs, debugging notes | memory/MEMORY.md or topic files | "vLLM START_DP_WAVE race condition" |
| Reusable workflow (multi-step, invoked on demand) | ~/.claude/skills/{name}/SKILL.md | /diagnose-bug, /upstream-bug-report |
| Code style / conventions (file-type scoped) | ~/.claude/rules/{name}.md | Python style, shell rules |
| Cross-project principles / persona / workflow rules | ~/.claude/CLAUDE.md | "Keep observations in sync" |
| Project architecture, design patterns, verification | {project}/CLAUDE.md | Architecture diagrams, test commands |
4. Write concisely
- Memory: facts, not narratives
- Skills: actionable steps, not essays
- Rules: short, scoped, triggered by file pattern
- CLAUDE.md: one bullet per rule
Anti-patterns
- Don't sediment session-specific context (current task, in-progress work)
- Don't duplicate what's already captured elsewhere
- Don't sediment unverified hypotheses — wait until confirmed