Key takeaways

• Memory failures are mostly manage-phase failures, and unlike stateless errors, they persist and compound.
• Drift (semantic, procedural, goal) comes from repeated lossy summarization; the defense is keeping raw records to reconcile against.
• Retrieval fails through semantic-versus-causal mismatch, memory blindness (the fact never resurfaces), and silent orchestration failures (no error, just slow decay).
• Validity fails through hallucinated facts and through staleness (correct once, now outdated); staleness is the most common.
• Self-reinforcing errors and over-generalization are where wrong memories become actively harmful.
• Safety failures include memory poisoning and cross-user privacy leakage; multi-agent topology affects how much leaks.
• These failures stem from real tensions (utility, efficiency, adaptivity, faithfulness, governance) that cannot all be maximized at once.

Continue to Chapter 5 for a framework that governs evolving memory to contain these failures.

---