Can agents fail from weak memory control rather than missing knowledge?

As multi-turn workflows grow, agents degrade — constraint focus is lost, errors accumulate, decisions drift. The instinct is to blame missing knowledge and add more context. This paper relocates the cause: the failures are driven by weak memory control, not knowledge gaps.

The two dominant patterns both fail in characteristic ways. Transcript replay appends prior interactions to the prompt, so context grows with turn count, attention selectivity drops, and early errors persist and reappear — hallucination carryover and constraint drift. Retrieval-based memory bounds prompt length but adds selection error: stale, conflicting, or injected artifacts perturb the current task state (the authors had to cap retrieval at three artifacts per turn to limit drift escalation).

The Agent Cognitive Compressor (ACC) replaces accumulation with a bounded, schema-governed internal state — the Compressed Cognitive State — updated by controlled replacement rather than growth. Crucially, it separates artifact recall from state commitment: the agent can consult artifacts without those artifacts automatically becoming persistent memory. This makes the write path explicit and auditable, and keeps the memory footprint bounded against poisoning and noisy recall.

The design echoes Is agent memory a storage problem or a connectivity problem? from the opposite direction: where that note argues retrieval quality is about access structure, ACC argues the more basic lever is gating what is allowed to commit at all. Both reject the "store everything, replay everything" default.