Agent AI Systems Lack a Memory Consolidation Phase, Not Just Capacity

Current AI agent systems rely on ever-larger context windows to handle memory limitations, but researchers argue this misses a fundamental design gap. Unlike biological brains, which use sleep-driven consolidation to transfer important short-term traces into durable long-term memory, AI agents have only a write path and a read path with nothing in between. Stored information is never revisited, merged, ranked by importance, or pruned of outdated facts after a task completes. Experiments tracking forced compactions in Claude Code found that seeded facts were lost entirely in the very first summary and never recovered in subsequent ones, suggesting capacity was not the issue. The missing element is a dedicated consolidation phase — an offline process that decides what mattered, generalises repeated observations, and structures memory rather than simply accumulating it.