Terms  /  Memory & State  /  Memory Bloat
03 · Memory & State

Memory Bloat

Accumulated stored context that is no longer relevant but still gets retrieved, increasing noise and cost. The memory-layer equivalent of Prompt Bloat.

Memory Bloat is what happens when a memory system stores everything and never cleans up. Old decisions that were reversed, temporary workarounds that were replaced, debugging notes from resolved issues, and experimental approaches that were abandoned — all sitting in memory, potentially being retrieved alongside current, relevant facts. Memory Bloat degrades Recall Precision and increases cost, because the system has to process and filter stale information on every retrieval.

Example
Over six months, an agent's memory has grown to 3,000 stored entries. An audit reveals that 40% are no longer relevant: references to a deprecated API, debug notes from bugs fixed months ago, and three different architecture decisions where only the latest one matters. Every retrieval query is searching through — and sometimes returning — stale entries. After cleanup, retrieval latency drops, relevance improves, and the model stops occasionally referencing the deprecated API in its suggestions. The memory got lighter and more useful at the same time.

More in Memory & State