# What is Forgetting?

Also called memory decay, memory expiry.

Forgetting is the deliberate removal or downweighting of stored memories so a system does not carry every past detail forever. Mechanisms include hard deletion on request, time-based expiry, relevance decay that lowers a record's retrieval weight as it ages, and supersession when a newer fact contradicts an older one.

Retention without limit degrades quality rather than improving it. Old preferences that no longer hold, resolved problems, and superseded documents keep matching queries and crowding out current information, so an assistant that remembers everything increasingly answers from a version of the world that no longer exists. Bounded memory is a correctness feature, not only a storage saving.

The mechanisms differ in reversibility and should not be confused. Decay reduces a record's influence while keeping it recoverable, expiry removes it after a defined period, and deletion is immediate and permanent. Systems typically decay by default, expire categories with known lifetimes such as session context, and reserve hard deletion for user requests and policy-mandated removal.

Deletion also has a legal dimension. Data protection regimes including the European GDPR and several United States state laws give individuals a right to have personal data erased, which for a memory system means erasing derived records too. A fact extracted from a deleted conversation, or an embedding computed from it, is still personal data, so deletion must reach the derived stores and any backups within the stated retention period.

The design risk runs in the opposite direction as well. Aggressive forgetting frustrates users who expect an assistant to remember what they said, and silent forgetting is worse than either extreme, because a person cannot tell whether a detail was never stored or was quietly dropped. Making retention rules visible, and letting a user pin a memory against decay, resolves most of this tension.

## Key points

- Bounded memory keeps answers current, not just storage small
- Decay, expiry, and deletion differ in reversibility
- Deletion must reach embeddings and extracted facts too
- Superseded facts should be demoted, not silently overwritten
- Retention rules should be visible and overridable by the user

## In practice

A memory store decays records not referenced in ninety days, so a note about a finished migration stops surfacing without being destroyed. When the user deletes their account, a job removes the conversation rows, the derived profile facts, and the embedding vectors computed from them, then confirms which stores were cleared and when the backups holding copies will roll off.

## Related terms

- [Long Term Memory](/en/glossary/long-term-memory)
- [Memory Consolidation](/en/glossary/memory-consolidation)
- [User Profile Memory](/en/glossary/user-profile-memory)
- [Episodic Memory](/en/glossary/episodic-memory)
- [Data Residency](/en/glossary/data-residency)

[Back to the AI Glossary](/en/glossary)