The moment an AI system remembers, it stops being only an interface. It begins to maintain a record of people, preferences, relationships, decisions and inferred patterns. That record may not look like a database. It may be a profile, summary, embedding, conversation history or provider-managed feature. Its effect is the same: future outputs change because the past was retained. Memory is therefore a records system whether the product calls it one or not.
Memory can become more authoritative than the source
Users often correct a document but forget the summary the model created from it. A relationship changes while the profile persists. An uncertain inference becomes a remembered fact. The AI repeats the memory confidently, giving it institutional weight. Derived memory needs provenance and an owner. Otherwise the system can no longer explain whether it remembers a source fact, a user statement or its own earlier interpretation.
Separate memory classes
- Session context: material required only for the current interaction.
- User preference: format, language and working style that can be changed safely.
- Operational state: current tasks, deadlines and decisions.
- Institutional fact: governed information sourced from authoritative records.
- Sensitive inference: conclusions about health, relationships, behaviour, risk or intent.
- Audit memory: logs retained to investigate use, error or abuse.
Each class requires different access, lifetime and correction. Sensitive inference should not become durable by default.
Value object — The AI Memory Ledger
For every persistent memory mechanism, record:
- Memory type and business purpose.
- Source and method of creation: direct, imported, inferred or summarised.
- Users and administrators who can read or alter it.
- Downstream models, tools and decisions it can influence.
- Retention, correction and deletion procedure.
- Permission changes that should trigger re-evaluation.
- Test evidence that deletion removes active and derived use.
Include vendor-controlled memory and local indexes. If the institution cannot inspect or delete a material memory, treat it as an external record.
Correction must propagate
Changing the source is insufficient if the old conclusion survives in embeddings, caches or user profiles. Design a correction event that identifies derived artefacts, invalidates them and prompts re-indexing. Preserve a limited audit record where necessary without allowing the obsolete fact to continue influencing answers.
Do not remember authority casually
The system should not persist claims such as “X can approve,” “Y is the owner” or “Z prefers this adviser” without authoritative source and effective date. Authority changes. Memory about authority can cause a direct control failure. Retrieve current mandate rather than relying on a conversational impression.
Make memory visible to the subject
Where appropriate, users should be able to see what the system has retained, why it was retained and how to correct it. The Swiss FDPIC emphasises transparency around purpose, functionality and data sources in AI-supported processing. Hidden memory creates hidden influence.
Forgetting is a capability
Institutions celebrate the model that remembers everything. Serious private systems should be designed to forget deliberately. The ability to expire, revoke and correct memory is what prevents convenience from becoming an uncontrolled institutional history.