One-sentence definition
Mechanism to preserve and retrieve relevant state between steps or sessions of a model-based system.
How to interpret this concept
In an assistant, memory may refer to the history of a conversation, a persistent state, or information retrieved between sessions. It does not necessarily mean the model learns by modifying its weights. The application decides what to store, for how long, for whom, and which part to reintroduce into the context.
Useful memory requires update and deletion rules. Storing everything indefinitely may retain sensitive data, obsolete instructions, and errors from previous conversations. Summarizing history reduces space but may lose nuances. Retrieved memories should be treated as data with provenance, not as highest-authority instructions.
Practical example: an assistant remembers that a person prefers brief responses. This preference does not authorize retaining banking data or sharing it with other users. Check isolation between accounts, expiration, correction of erroneous memories, and effective deletion. Also distinguish remembering a preference from remembering external facts that may have changed.
