Stale Memory Is Worse Than No Memory: Why AI Agents Need to Forget
Ask a team what agent memory should do and everyone says "remember." Watch the same team a month after wiring it up and the complaint has inverted: the agent remembers too well. It cites the state-management library you migrated off in June, follows a refactoring guideline retired last quarter, and treats an abandoned experiment as house style. An agent that forgets nothing eventually knows everything except what is currently true.
TL;DR
- Similarity search has no concept of stale: a deprecated convention embeds exactly as well as a current one, forever.
- Forgetting is not one operation but three: close facts that were superseded, decay what goes unused, suppress what misleads.
- Deletion is the wrong default. A closed fact with a validity window still explains past decisions; a deleted one leaves a hole where an explanation should be.
Why stores never forget on their own
A vector store ranks by closeness in embedding space. Age, deprecation, and supersession are not dimensions of that space. "Use the ORM's lazy loading" written in January is retrieved in September with unchanged confidence, because nothing about September appears in the query vector. The store is not broken; it answers a question ("what is most similar?") that is different from the one the agent needs answered ("what is currently true?").
Raw accumulation makes this worse in a specific way: the more diligently your agents write memory, the more retired truths the store contains, and the higher the odds any given recall surfaces one. Teams that celebrate their growing memory count are often measuring their future confusion.
Forgetting is three operations, not one
Closing. When a fact is superseded, do not overwrite or delete it. Mark it closed, record its validity window, and link its successor: "we use npm, true until 2026-06, superseded by pnpm." Recall prefers active facts, but the closed record still exists, and that matters more than it seems: the agent asked to explain June's lockfile needs June's truth. Overwriting rewrites history; closing versions it.
Decay. Entries that are never recalled, never confirmed, and never contradicted should not keep competing at full strength forever. Letting unused associations weaken over time moves them down the ranking without destroying them. The failure mode to avoid is silent destruction: decay that quietly deletes is indistinguishable from data loss. Decay should demote, not erase.
Suppression. Some memories are not old, just wrong, and known to be wrong because acting on them failed. Those deserve an explicit negative mark that pushes them below the alternatives (the mechanics of that outcome channel are the subject of why agents repeat the same mistakes). Suppression is targeted; decay is ambient.
Hard deletion keeps a narrow, real role: secrets that should never have been stored, personal data that must go on request, and genuine noise with no decision history behind it. The test is simple: if nothing ever depended on the entry, delete freely; if anything did, close it instead.
What this looks like in practice
With plain files, forgetting is an editorial habit: a status line on each entry (active / closed, superseded by…), a periodic pruning pass that produces a reviewable list instead of silent removals, and the discipline never to rewrite an old entry in place. It works, and its cost is that a human is the garbage collector.
With a memory system, the same three operations become properties to look for: statuses and validity windows on facts, use-dependent decay that demotes rather than deletes, and a feedback channel for explicit suppression. An offline consolidation phase (the subject of why agent memory needs sleep) is where much of this naturally runs: merging duplicates, closing the superseded, and letting the unused sink, between sessions rather than during them.
The uncomfortable summary: "never forgets" is a storage property advertised as a virtue. In a working agent, it is a bug with good marketing. Current beats complete.
Related: The AI Agent Memory Crisis on why bigger context windows make staleness worse, not better. Agent Memory: Evidence versus Policy on the authority model behind closing facts.