# 6-Tier Memory Architecture — Quick Reference

> AI Starter Package (aistarterpackage.com)

## The 6 Tiers

| Tier | File/Location | Scope | Retention | Max Size |
|------|--------------|-------|-----------|----------|
| 1. Session | Context window | Current conversation | Until session ends | ~200K tokens |
| 2. Working | `memory.md` | Active task context | Pruned weekly | 100 lines |
| 3. Knowledge | `knowledge-base.md` | Learned rules | Permanent (auditor-gated) | 200 lines |
| 4. Agent | `.claude/agent-memory/` | Per-agent knowledge | Permanent | Per-agent file |
| 5. Graph | MCP knowledge graph | Structured entities | Permanent | Unlimited |
| 6. Archive | `memory/YYYY-MM-DD.md` | Daily session logs | Permanent | One file/day |

## When to Use Each Tier

- **Need it RIGHT NOW?** → Session memory (context window)
- **Need it THIS WEEK?** → Working memory (memory.md)
- **Need it FOREVER?** → Knowledge base (auditor reviews first)
- **Agent-specific?** → Agent memory (each agent has its own file)
- **Structured relationships?** → Knowledge graph (entities + edges)
- **What happened on a date?** → Daily archive (chronological logs)

## Maintenance Rituals

| Command | When | What It Does |
|---------|------|-------------|
| `/morning` | Start of day | Load context, show priorities, check knowledge graph |
| `/sync` | Mid-day | Refresh working memory, process nominations |
| `/wrap-up` | End of day | Archive learnings, update memory, plan tomorrow |
| `/safe-clear` | When context is heavy | Save state, flush context, resume fresh |

## Memory Hygiene Rules

1. Keep `memory.md` under 100 lines — prune aggressively
2. Keep `knowledge-base.md` under 200 lines — quality over quantity
3. Every knowledge-base entry needs `[Source:]` provenance
4. Run `/sync` after 30+ tool calls to prevent context degradation
5. Archive completed tasks during `/wrap-up` — don't carry stale items

## Self-Improvement Loop

```
Observation → Nomination → Auditor Review → Knowledge Base Promotion
```

Not every learning becomes permanent. The Auditor agent reviews nominations
before they're promoted to the knowledge base. This prevents bloat and
ensures high signal-to-noise.

---

*Get the full 6-tier memory system pre-configured at aistarterpackage.com*