Agent Memory Paper Says Retrieval Isn’t Learning

Sharp thesis, and it lands where agent builders feel the pain: more context and better retrieval can improve recall, but they do not create lasting competence.

• –The paper draws a clean line between replaying stored examples and learning abstract rules, which reframes “memory” as an optimization problem, not a UX feature.
• –Its strongest practical warning is memory poisoning: if agents keep reusing injected notes across sessions, bad data can persist far beyond a single interaction.
• –The Complementary Learning Systems framing is useful because it suggests a hybrid stack, not a false choice between RAG and training.
• –For builders, the implication is simple: persistent context helps coordination, but if you want real capability gains over time, you need some form of consolidation or fine-tuning.
• –There are no empirical benchmarks here, so this is mainly a conceptual paper, but it sets a useful standard for evaluating agent memory claims.