Is agent memory capacity or quality the real bottleneck?

The intuitive picture of agent memory is a storage problem: give the agent more room and it remembers more. The system-scaling analysis rejects this. "Memory is not merely a storage layer; the harder problem is memory quality" — what to store, what to discard, how to retrieve the right information at the right time, and how to avoid staleness, drift, contamination, and over-generalization. Adding capacity without curation makes things worse: more stored material means more stale entries to retrieve, more opportunities for contaminated content to surface, and more room for over-generalized lessons to misfire on cases they do not fit.

This reframes memory engineering as a discarding and curation problem, not an accumulation one. The failure modes are specific and diagnosable — staleness (kept too long), drift (slowly diverging from ground truth), contamination (bad entries poisoning retrieval), over-generalization (a narrow lesson applied too broadly) — and each calls for different hygiene. The open question, which the paper leaves unresolved, is what discarding policy avoids all four without throwing away genuinely useful long-tail knowledge. The counterpoint is that aggressive forgetting risks losing rare-but-critical information, so quality is a trade-off, not a free win. This matters because it redirects effort from bigger memory stores toward better forgetting — the part of memory design that is hardest and least solved.

— "From Model Scaling to System Scaling: Scaling the Harness in Agentic AI", https://arxiv.org/abs/2605.26112