Memory-augmented LLM agents were pitched as a way to dodge catastrophic forgetting entirely — don't retrain, just remember. This paper runs the classic continual-learning protocol on that setup and finds the same bottleneck resurfaces, relocated from parameter capacity to retrieval capacity: retrieval pollution, context competition, and memory dilution.
Parametric CL overwrites weights; Memory CL overwrites nothing — it just runs out of room to retrieve. Hover the bottleneck boxes.
The paper's core claim, as data: the carrier and interference mechanism change, the shape of the problem does not.
A BM25 query pulls back top-k memory entries — some relevant to the current task, most not. Hover a slot.
Toggle memory-store size. As more entries compete for the same fixed context window, useful memories get displaced.
Heatmap: rows are memory slots, columns are time steps as the pool grows. Color = probability a query for old content still surfaces this slot. Hover any cell.
Detail that helps within one task actively misleads a different task. Toggle environment.
Raw's detail barely hurts easy cases — it wrecks the hard ones.
Insight cuts forgetting too, not just forward transfer.
Toggle direction. The condition that adapts best to the new task can damage the old one most, in the same run.
Cond-Ind's fine-grained index only helps if the underlying content is actually diverse. On repetitive content, near-identical entries funnel every query into the same handful of items.
Every memory design is a value (how experience is written) crossed with a key (how it's indexed). Hover a cell for its definition; highlighted cells are the ones actually tested.