feat(memory): add MemoryCapability with SQLite backend

[mustafabozkaya]

Summary

Add persistent memory capability for Pydantic AI agents. This is the first implementation of the Persistent key-value memory feature request.

What This Adds

New Capability: MemoryCapability

Provides five tools for persistent memory management:

• memory_store — Save information with key, value, tags, and metadata
• memory_retrieve — Full-text search across stored memories (FTS5)
• memory_list — List entries with glob pattern and tag filtering
• memory_delete — Remove an entry by key
• memory_compact — Clean up old, rarely-accessed entries

Architecture

The implementation follows the established CodeMode capability pattern:

• _abstract.py — AbstractMemoryBackend interface for custom backends
• _capability.py — MemoryCapability (AbstractCapability subclass)
• _models.py — Pydantic models (MemoryEntry, input schemas)
• _sqlite.py — SQLiteMemoryBackend (default backend with FTS5)
• _toolset.py — MemoryToolset (tool implementations)

Key Features

• SQLite backend with FTS5 full-text search
• AbstractMemoryBackend interface for custom storage engines
• Tag-based filtering and glob pattern matching
• Access tracking with automatic compaction
• Lazy initialization of database connection
• 18 tests covering models, backend, and edge cases

Usage

from pydantic_ai import Agent
from pydantic_ai_harness import MemoryCapability

agent = Agent('openai:gpt-5', capabilities=[MemoryCapability()])
result = agent.run_sync('Remember that my favorite color is blue')
result = agent.run_sync('What is my favorite color?')  # blue

Checklist

• Follows CodeMode capability pattern
• Ruff lint clean
• 18 tests passing
• README.md with examples
• Abstract interface for extensibility

Closes #179