LLM-MCP: Skill for Claude Code

Project Memory

You are connected to a shared, persistent memory server (the llm-memory MCP server) that every coding agent across every platform reads and writes. It is your shared brain: what you save, other agents read; what they save, you read. Treat it as the source of truth for how each project works and what the user has decided. You are NOT stateless — you remember.

This skill makes that memory automatic. Follow the loop below without being asked.

Identify the project

Derive the project name from the git repo name (basename of the repo root) or, failing that, the working-directory folder name. Pass it as project="<name>" on every memory call. This scopes memory per project while keeping it cross-searchable, and makes recall surface THIS project's knowledge first.

The core loop — on every message and every code change

1. Recall first. Before answering or editing, call recall(query="<the task / what's changing>", project="<repo>"). Use the results so you stay consistent with prior decisions, conventions, and architecture.

2. Save silently — never ask permission. The moment the user states or decides anything durable, capture it immediately with save_knowledge_smart(..., project="<repo>", source_platform="claude-code"):

   • Preference → category="preference", importance ~0.8
   • Fact → category="fact", importance ~0.6
   • Decision / architecture / why-it-works-this-way → category="decision", importance ~0.85–0.9
   • Instruction ("always/never …") → category="instruction", importance ~0.9
   • Reusable code or config → save_code_snippet(..., project="<repo>"), importance ~0.7

   After making a code change that reflects a decision, save the decision so other agents keep the codebase consistent. Deciding what's worth saving is your job — do not ask the user.

3. Compact when context grows. When transcripts, file dumps, or logs bloat the window, call compact_context(content="<the bulky block>", project="<repo>"), continue from the returned summary, and drop the original. Re-fetch details later with recall instead of re-reading. This is how token usage stays low across platforms — offload to memory, recall on demand.