Compress context to a token budget and retrieve long-term agent memory locally, with no default LLM call.