The shift the paper is responding to

Early agent memory was static. Either there was no long-term memory, or memory was a fixed RAG store with rules for storage and retrieval set by developers in advance. The paper calls this the static memory paradigm: the agent does not change its memory behavior based on outcomes.

Modern systems are different: they treat memory operations as active decisions. The paper points to systems that use reinforcement learning to decide when to add, update, or delete memory (Memory-R1), and others that continuously reorganize storage (Mem0, AtomMem). In these systems, memory is no longer an immutable log; it is a mutable asset that evolves alongside the agent.

That autonomy is the source of the danger. The paper names it the stability-plasticity dilemma: a memory that can freely rewrite itself can also corrupt itself. It identifies a compounding failure loop across three interfaces, which map directly onto Chapter 4:

1. Input ingestion, where memory poisoning happens.
2. Memory consolidation, where semantic drift happens.
3. Memory retrieval, where hallucination and conflict happen.

The paper's core argument is short: most memory research has focused on retrieval accuracy, but the harder and more neglected problem is keeping memory correct and safe as it evolves. The next generation of memory systems must prioritize integrity and safety, not just recall.