Optimize prompts and tool calls to stay within token limits.
Large prompts and tool messages exceed limits, forcing truncation of earlier turns.
Granular long‑term memory with namespaces, TTL, and export.