AI memory fails in three distinct ways — and piling on more storage makes all three worse, not better.

This question separates three failure modes that often get lumped together as 'memory degrades over time.' The corpus suggests they're mechanically distinct. The clearest single statement comes from work arguing the real bottleneck is quality, not capacity: piling up more stored material without curation increases staleness (facts that were true once but aren't now), contamination (bad or irrelevant entries that pollute retrieval), and over-generalization all at once—and crucially, more storage makes performance *worse*, not better What makes agent memory quality better than storage capacity?. So the three aren't just labels; they're separate ways a growing store turns against the agent.

Drift shows up most concretely as something that *accumulates* rather than something that's simply outdated. One study of continuously consolidated memory found an inverted-U curve: early consolidation helps, but as experience piles up the memory eventually performs worse than just keeping raw episodes—one model failed 54% of problems it had previously solved. The named mechanisms are revealing: misgrouping (unrelated experiences merged), applicability stripping (the conditions that made a lesson valid get filed off), and overfitting to narrow streams Does agent memory degrade when continuously consolidated?. That's drift as a *processing* artifact—the act of compressing and summarizing introduces error. It's different from staleness, where the entry was recorded faithfully but the world moved.

Drift also appears in long workflows as constraint drift: agents lose track of requirements not because they lack knowledge but because transcript replay and retrieval lack any gate on what gets written or trusted. The proposed fix is a bounded, schema-governed committed state that separates temporary artifact recall from permanent memory writes—so errors can't silently compound into the long-term store Can agents fail from weak memory control rather than missing knowledge?. This points at the deeper pattern: contamination and drift are both write-side problems, which is why several systems attack them at the write/management boundary. One framework splits memory into an explicit hot-path (the agent deliberately decides what to keep via tool calls) and an implicit background path (programmatic triggers), trading context-sensitivity against reliability for generation, storage, retrieval, *and* deletion How should agents decide what memories to keep?.

The corrective theme across the collection is that good memory must actively prune, not just accumulate. FluxMem treats memory topology as something that continuously forms, refines, and removes links based on execution feedback—explicitly to eliminate the interference that fixed, ever-growing retrieval suffers Should agent memory adapt dynamically based on execution feedback?. Autonomous memory folding consolidates history into structured episodic, working, and tool schemas, and the authors are careful to note that *structure plus agent autonomy* is what avoids the degradation that plagues naive consolidation Can agents compress their own memory without losing critical details?—a direct contrast with the unguarded consolidation that produced the inverted-U collapse above.

The quietly useful takeaway: the cure differs by failure mode. Staleness wants deletion and recency-aware curation; drift wants gating and applicability-preserving consolidation (or restraint about consolidating at all); contamination wants write-side admission control. A system tuned to fight one can be wide open to the others—which is why the corpus keeps returning to *what to discard and when* as the governing question rather than how much to store What makes agent memory quality better than storage capacity?.