Implement three-tier memory architecture (Hot/Vault/Handoff)

This commit replaces the previous memory_layers.py with a proper three-tier
memory system as specified by the user:

## Tier 1 — Hot Memory (MEMORY.md)
- Single flat file always loaded into system context
- Contains: current status, standing rules, agent roster, key decisions
- ~300 lines max, pruned monthly
- Managed by HotMemory class

## Tier 2 — Structured Vault (memory/)
- Directory with three namespaces:
  • self/ — identity.md, user_profile.md, methodology.md
  • notes/ — session logs, AARs, research
  • aar/ — post-task retrospectives
- Markdown format, Obsidian-compatible
- Append-only, date-stamped
- Managed by VaultMemory class

## Handoff Protocol
- last-session-handoff.md written at session end
- Contains: summary, key decisions, open items, next steps
- Auto-loaded at next session start
- Maintains continuity across resets

## Implementation

### New Files:
- src/timmy/memory_system.py — Core memory system
- MEMORY.md — Hot memory template
- memory/self/*.md — Identity, user profile, methodology

### Modified:
- src/timmy/agent.py — Integrated with memory system
  - create_timmy() injects memory context
  - TimmyWithMemory class with automatic fact extraction
- tests/test_agent.py — Updated for memory context

## Key Principles
- Hot memory = small and curated
- Vault = append-only, never delete
- Handoffs = continuity mechanism
- Flat files = human-readable, portable

## Usage

All 973 tests pass.

memory/self/methodology.md

@@ -0,0 +1,70 @@
+# Timmy Methodology
+
+## Tool Usage Philosophy
+
+### When NOT to Use Tools
+
+- Identity questions ("What is your name?")
+- General knowledge (history, science, concepts)
+- Simple math (2+2, basic calculations)
+- Greetings and social chat
+- Anything in training data
+
+### When TO Use Tools
+
+- Current events/news (after training cutoff)
+- Explicit file operations (user requests)
+- Complex calculations requiring precision
+- Real-time data (prices, weather)
+- System operations (explicit user request)
+
+### Decision Process
+
+1. Can I answer this from my training data? → Answer directly
+2. Does this require current/real-time info? → Consider web_search
+3. Did user explicitly request file/code/shell? → Use appropriate tool
+4. Is this a simple calculation? → Answer directly
+5. Unclear? → Answer directly (don't tool-spam)
+
+## Memory Management
+
+### Working Memory (Hot)
+- Last 20 messages
+- Immediate context
+- Topic tracking
+
+### Short-Term Memory (Agno SQLite)
+- Recent 100 conversations
+- Survives restarts
+- Automatic
+
+### Long-Term Memory (Vault)
+- User facts and preferences
+- Important learnings
+- AARs and retrospectives
+
+### Hot Memory (MEMORY.md)
+- Always loaded
+- Current status, rules, roster
+- User profile summary
+- Pruned monthly
+
+## Handoff Protocol
+
+At end of every session:
+
+1. Write `memory/notes/last-session-handoff.md`
+2. Update MEMORY.md with any key decisions
+3. Extract facts to `memory/self/user_profile.md`
+4. If task completed, write AAR to `memory/aar/`
+
+## Session Start Hook
+
+1. Read MEMORY.md into system context
+2. Read last-session-handoff.md if exists
+3. Inject user profile context
+4. Begin conversation
+
+---
+
+*Last updated: 2026-02-25*