[ { "prompt": "What is the safest design rule for final answers?", "answer": "The safest design rule is: the latest `USER TASK` is the authority, and everything else is only bounded evidence. Scaffolds, routes, memory, profiles, and tool traces should never override the user's current request or appear in the final answer unless the user asks to see them.", "skills": [ "runtime_debugging" ], "route": "runtime_debugging", "judge_used": false, "uncertain": false, "fallback_reason": "", "generation_meta": {}, "summary": "", "flush_count": 0 }, { "prompt": "Add a note that old runtime context should not override my latest user prompt.", "answer": "Old runtime context should never override the latest user prompt. The runtime should keep old context in a compact memory summary, retrieve only the parts that directly help, and make the final composer answer the current `USER TASK` first.", "skills": [ "runtime_debugging" ], "route": "runtime_debugging", "judge_used": false, "uncertain": false, "fallback_reason": "", "generation_meta": {}, "summary": "", "flush_count": 0 }, { "prompt": "Give me a short checklist for this agentic session memory system.", "answer": "Use this agentic session checklist:\n1. Keep the latest user prompt in a dedicated `USER TASK` field.\n2. Store older turns in rolling memory summaries instead of appending everything forever.\n3. Keep tool results, route decisions, and profile state as artifacts, not prompt noise.\n4. Flush old turns when the live window gets too large.\n5. Use KV cache only for the active generation window; use summaries for long-horizon continuity.\n6. Regression-test topic switches so old memory cannot hijack the new task.", "skills": [ "runtime_debugging", "agi_architecture" ], "route": "runtime_debugging", "judge_used": false, "uncertain": false, "fallback_reason": "", "generation_meta": {}, "summary": "", "flush_count": 0 } ]