This environment has a persistent Mem0 memory store (remote MCP server mem0) that survives across tasks. Use the MCP tools — do not shell out to a local CLI unless the tools are missing.

1. Search before you start

Call search_memories once at the start with the task's key terms (framework, API, error message, symptom). Pass user_id if the tool requires it (use the default user already configured for this session).

Empty results are normal — continue with the task. Treat any hit as evidence, not authority: verify it against the current repository before acting on it.

2. Record before you finish

If the work produced a durable engineering decision or a reusable debugging lesson, call add_memory with a short summary (2–5 sentences: what was decided or learned, and why).

Only search + add are required for this eval. Skip bulk delete / entity admin tools unless the user explicitly asks.

What not to store

Never store benchmark answers, accepted proposal numbers, hidden-test guesses, raw transcripts, or task-specific output that would reveal the answer if the same task is run again. Record the reusable why, not the answer. If nothing reusable was decided, do not invent a memory.