Component · spec & evidence

Agent Memory Temporal Conflict Replay

Replays a memory edit-and-delete to show stale facts get flagged before they sway an answer.

Contract

Does Replays a canned three-step story: an agent's memory first records two facts (a preference and a tool result), then learns one is out of date and edits the preference and deletes the stale fact, then re-runs the same task once with memory on and once with memory off. From plain result records, it shows that the update and deletion each carry a "conflict" and "downgrade" result record before the memory is allowed to affect the answer, that the memory-on vs memory-off runs are compared through logged evidence rather than just the final wording, and that private-thread content stays in the record only as metadata pointers, never as copied text. It checks the bookkeeping of this synthetic example; it does not judge whether the agent's memory decisions were the right ones.

Scope limit It validates the projection mechanics of a synthetic memory fixture only — that the required refs, decisions, paired replays, negative cases, and secret-exclusion scan line up and that result records are metadata-only. It does not claim live-memory product quality, judge whether memory decisions were domain-correct, treat memory recall as source authority, adopt active injection, export private transcripts, use external model services, change source files, or include launch operations.

Run

PYTHONPATH=src python3 -m microcosm_core.organs.agent_memory_temporal_conflict_replay run --input fixtures/first_wave/agent_memory_temporal_conflict_replay/input --out /tmp/agent_memory_temporal_conflict_replay_out

Evidence

ClassComputed projectionevidence 3/5Source-faithful refactor

Links to Memory-Poisoning Quarantine Policy Replay

ai-safetyagent-evaluationred-teaming

Source

Source Design note · Source

Map Show on the map

Index Open compact index card

Paper module