Memory or Fine-Tuning for Personalization? Start with What Must Change

Use retrieved memory for user-specific facts that change and need inspection, correction, or removal. Consider fine-tuning when the desired change is a repeatable model behavior supported by a suitable training and evaluation process. These approaches can coexist; they solve different parts of personalization.

A user's current project, preferred response format, and account state do not automatically justify a separate trained model. Start by identifying the information or behavior that the base application fails to provide.

Distinguish facts from behavior

A changed delivery address is a current fact and belongs in the authoritative application system. A preference for short summaries can be supplied as explicit context. A consistent domain-specific output pattern may be a candidate for prompt design or training, depending on the model and workload.

Do not use memory as a substitute for enforcing a strict output schema. Conversely, do not assume a training update provides a convenient way to inspect and remove one person's current preference.

Build the simplest comparison first

Evaluate the existing prompt, the prompt plus relevant user context, and any proposed training approach on the same tasks. Keep source facts and user populations separated between training and evaluation where applicable.

Measure task completion and correction effort, not only whether the answer sounds personalized. An assistant that repeatedly mentions remembered details may feel intrusive while contributing nothing to the task.

Give changing information a lifecycle

Record where a preference came from and whether it was explicit or inferred. Test an exception, a permanent correction, and a removal request. Current permissions must remain authoritative regardless of what prior interactions suggest.

For retrieval-based personalization, inspect the exact facts supplied to the model. A failure may come from a missing record, the wrong scope, or poor application of correct evidence. Those require different fixes.

Choose from operational requirements

Ask how often the relevant information changes, who can correct it, how its use is explained, and how a failure is reversed. Include the cost of evaluations and maintenance, not just one training run or one retrieval request.

The user-profile guide describes a small inspectable starting point. Supermemory's profile documentation covers its user-context path. To test whether memory solves the actual gap, start a Supermemory pilot with one preference and one real task before considering a broader model change.