Forgetful adds procedural, prospective memory to agent MCP

Forgetful is evolving from a simple RAG store into a unified cognitive architecture for multi-agent systems by abstracting planning and skills into the memory layer.

• –v0.3.0 shifts the focus from "what the agent knows" to "what the agent can do" and "what it is trying to achieve" through procedural and prospective memory.
• –Storing skills as first-class citizens allows for cross-platform agent capabilities, enabling behavior sharing between diverse tools like Claude Code and Gemini CLI.
• –The prospective memory feature provides a shared source of truth for long-term objectives, solving the complex coordination problem in multi-agent workflows.
• –A "meta-tool" pattern balances deep functionality (over 40 sub-tools) with strict token budget management to avoid context window overflows.
• –Its commitment to local-first performance using SQLite and FastEmbed makes it a strong, privacy-centric choice for the local LLM community.