Why it matters
Wojciech Pluta’s recorded mortgage-agent example preserves an output-format preference across sessions, rejects a blanket approval instruction and sends a proposed shared practice for human review. His memory pipeline classifies candidates, checks evidence and existing records, and lets the harness decide whether to persist them. These prepared cases illustrate admission controls, not comprehensive resistance to memory poisoning.
My takeaway: Give each memory type explicit evidence, ownership and scope requirements. Test a harmless preference, a forbidden policy override, a conflicting record and a proposal to share memory across users. Restart the session to inspect what actually persists and verify access boundaries during retrieval. Record provenance and supersession decisions, and ensure an agent cannot bypass the admission path with a direct write.