ngocsangyem Memory System Toolkit
Reference guides for the memory system: session capture, pattern extraction, and cost tracking.
When to Use
- During Phase 0 (Orient) to load previous session context
- During Phase 6 (Reflect) to capture learnings and patterns
- When the
analystagent tracks costs or extracts patterns
Workflow Integration
Operates in Phase 0 (Orient) and Phase 6 (Reflect). Output supports the analyst agent.
References
| Reference | When to load | Content |
|---|---|---|
| capture-architecture.md | Before any agent-side memory write | The 2-path contract — ##prefix: is user-typed only; agents write via direct Edit |
| session-capture.md | Phase 6 | Capturing session learnings in 3 categories (patterns/decisions/failures) |
| pattern-extraction.md | Phase 6 | Extracting high-frequency patterns for instruction-file promotion |
| cost-tracking.md | Phase 0, 6 | Token usage tracking, cost reporting, budget alerts |
| consolidation.md | Manual | Prune stale entries, merge duplicates, archive cost data (run when memory grows large) |
How session-capture is invoked
session-capture is not a CLI subcommand. At Phase 6 (Reflect) the agent reads references/session-capture.md and follows its 4 steps using Read / Write / Edit directly. There is no mewkit memory session-capture script — content extraction requires LLM analysis that a static CLI cannot produce.
Subcommands
--prune
Prune old standard-severity entries from canonical JSON stores, then regenerate Markdown views.
Stores subject to pruning: fixes.json, review-patterns.json, and
architecture-decisions.json.
What gets pruned:
## headingswith a date(YYYY-MM-DD, severity: standard)older than threshold (default: 90 days)
What is exempt:
severity: criticalorseverity: securityentries — never pruned- Entries without a parseable date — kept as-is (safe default)
How to run:
the memory skill --prune # default 90-day threshold
the memory skill --prune --days 180 # custom threshold
the memory skill --prune --dry-run # show what would be archived without moving
Mechanism: prune schema entries in JSON by lastSeen and severity, then run
mewkit memory render-views. Report the affected stores and entries.
Recovery: Copy entries from lessons-archive.md back to the appropriate topic file.
Schema Notes
Split JSON files (fixes.json, review-patterns.json, architecture-decisions.json) all use v2.0.0 schema with fields: version, scope, consumer, patterns[], metadata. Each pattern entry supports: id, type, category, severity, domain, applicable_when, context, pattern, frequency, lastSeen.
Gotchas
- Stale patterns applied to changed codebase: Memory suggests patterns from old architecture → Run consolidation when patterns exceed 50 entries; flag patterns with lastSeen > 6 months
- cost-log.json growing unbounded: Every session appends without pruning → Run consolidation to archive entries older than 90 days
- Generated views are not writable: Do not write
lessons.mdor Markdown topic files. Update the canonical JSON store, then render views.