Problem

Current tool descriptions consume ~12K tokens (30 tools x ~400 tokens). On a 32K context model, that is 37% consumed by tool descriptions alone. Static compression is a hack.

Approach

Dynamically load tools based on conversation context using the memory system. Memory informs which tools to activate per session. Tools not relevant to the current conversation are not loaded into the prompt.

Design:

• Tool registry supports lazy loading — tools registered but not included in prompt by default
• Memory system tracks which tools the user frequently uses
• Conversation context (topic, keywords) determines which tool categories to activate
• Core tools (read_file, search, RAG) always loaded
• Specialized tools (Playwright, MCP servers, shell) loaded on demand
• Tool set can change mid-conversation as context evolves

Dependencies

• MemoryStore: persistent memory data layer (SQLite + FTS5) #542 (MemoryStore) — tracks tool usage patterns
• MemoryMixin: 5 memory tools + 3 lifecycle hooks for any agent #543 (MemoryMixin) — provides tool selection context

Acceptance Criteria

• Tool prompt reduced from ~12K to ~3-4K tokens in typical sessions
• Relevant tools activated automatically based on conversation context
• Core tools always available regardless of context
• No degradation in tool selection accuracy
• Memory tracks which tools user uses most frequently