feat(memory): Add three-layer memory system with decision routing

[AmariahAK]

PR Description

i. What was the issue

Mini-Agent had a flat, single-layer memory system (SessionNoteTool + RecallNoteTool with a simple JSON store) and no structured decision routing. This limited the agent's ability to:

• Maintain immutable project rules separate from learned patterns
• Retain cross-session learnings with confidence gating
• Auto-cleanup session-scoped context
• Classify and route tasks by complexity (simple → direct, complex → think→plan→execute→verify→reflect→learn, conflict → confidence-weighted dispatch)

Issue #110 proposed a HeartFlow-inspired three-layer memory architecture and decision routing rules.

ii. Where was the issue

Layer	Before	After
mini_agent/tools/note_tool.py	Flat JSON record/recall, no confidence	Unchanged — remains as-is for backward compat
System prompt	No memory guidance, no routing rules	53 new lines: Memory System + Decision Routing
Config	No memory config	MemoryConfig with enabled, core_memory_path, learned_confidence_threshold
CLI tool init	No memory tools	Three new tools registered alongside existing ones

iii. How did you fix it

New module mini_agent/memory/ — 5 files, all following existing architecture:

• schema.py: MemoryEntry (Pydantic, confidence 0.0-1.0) + MemoryStore (JSON persistence, follows SessionNoteTool pattern)
• manager.py: MemoryManager — CORE (read-only JSON), LEARNED (confidence-gated writes ≥0.7 threshold), EPHEMERAL (in-memory, reset() on session end)
• tools.py: CoreMemoryTool, LearnedMemoryTool, EphemeralMemoryTool — all extend the existing Tool base class with name/description/parameters/execute → ToolResult

Config: MemoryConfig model added to ToolsConfig, parsed in from_yaml() following the MCPConfig pattern.

System prompt: New ## Memory System section (3 layers documented) + ## Decision Routing section (simple/complex/conflict routing with 6-step think→plan→execute→verify→reflect→learn cycle).

CLI: Memory tools registered in initialize_base_tools() with the existing green-checkmark pattern. EPHEMERAL reset called at session start and end.

iv. Why did you choose that method over others

• Tools extend Tool base — zero changes to the Agent.run() loop. Tools are registered by name into self.tools dict, dispatched by the LLM automatically. No new infrastructure needed.
• JSON storage — consistent with the existing SessionNoteTool pattern (load/save via MemoryStore.load_from_file / save_to_file). No new dependencies.
• Config-driven — follows MCPConfig nested-model pattern with defaults, parsed in from_yaml(). Users can disable memory or tune thresholds without touching code.
• Confidence gating at manager level — the MemoryManager.record_learned() returns (accepted: bool, reason: str), keeping the tool layer thin and the gating logic centralized.

v. How to test it locally

# Run the full memory test suite
uv run pytest tests/test_memory.py -v
# 19 tests: schema, manager, tool layers, confidence gating, ephemeral reset

# Verify config loads with memory defaults
uv run python -c "from mini_agent.config import Config; c = Config(api_key='test', model='test'); print(c.tools.memory)"
# Expected: MemoryConfig(enabled=True, core_memory_path='', learned_confidence_threshold=0.7)

# Verify memory tools appear in CLI (requires API key in config.yaml):
# uv run python -m mini_agent.cli
# Look for: ✅ Loaded Core Memory tool / Learned Memory tool / Ephemeral Memory tool

# Run full non-integration test suite
uv run pytest tests/ --ignore=tests/test_agent.py --ignore=tests/test_integration.py --ignore=tests/test_acp.py --ignore=tests/test_llm_clients.py --ignore=tests/test_llm.py --ignore=tests/test_session_integration.py -v