GLOSSARY
Memory
Memory in an AI application means stored information that becomes context for later interactions. It can include confirmed preferences, previous decisions or unresolved questions. The implementation determines what is retained, where it is available and how long it remains relevant.
Continuity without carrying assumptions forward
The last agreed price can be useful when renewing a contract. A suspected budget constraint from that period should remain identifiable as an earlier assumption. Treating both statements as equally current can distort the team's view of its negotiating room.
Useful memory therefore needs provenance, timestamps and a correction process. Owners should decide which information may be shared and which belongs to an individual or a particular case. This can reduce repeated research without turning every conversation note into permanent organisational knowledge.
Assess reuse and storage together
Evidence starts with recurring tasks: which questions no longer need asking, which stored statements prove inaccurate, and how quickly corrections take effect. A saved preference also needs a clear owner when colleagues disagree about its meaning. Retention in logs and backups requires separate attention. Saving a conversation does not automatically update a model's parameters; that is the distinct process described as training.