Memory
Information an AI product saves from earlier work and gives the AI again later. It does not change the model.
Memory is information an AI product saves from earlier work and adds to later work. A product may save it under a staff account or share it with a team. The firm should check whether it keeps details separate for each client. When it gives a saved memory to the AI, that memory becomes part of the context for the new task. It does not change what the underlying model learned in training.
On Monday, a mortgage adviser asks the firm's AI product to draft a gentle email to Leanne Price, whose house purchase is paused after she was made redundant. The product saves a memory under the adviser's account: "The client has been made redundant; use a gentle tone." On Wednesday, the adviser asks for an update to Mark Ellis about his remortgage application. The product adds the saved line to that request, and the draft mentions Mark's redundancy. Mark has been in the same job for ten years.
The problem is where this product stored the memory. A file note belongs to Leanne's client record; the saved line was attached to the adviser and reached Mark's work too. The adviser checks the draft before sending it. The firm also needs to check whose later work a product's memory can enter, what it saves without being asked, and how to remove a mistaken memory.
Not to be confused with
- The context is the material the AI has for one piece of work. A memory affects the work only when the product adds it to the context.
- A product may include earlier messages in the context of a new request within the same conversation. That is the current conversation's context; memory is information the product saves separately for reuse. Check what the product carries into a new conversation.
- Training is how the company that built the model made it, and it is not staff training or CPD. Saving a memory leaves the model unchanged. Whether that company uses the firm's conversations to train later models is a separate question. Check the product's settings and vendor terms, then ask the vendor if they leave it unanswered.
- Model, in this entry, means the part of an AI product that writes, never a model portfolio.
Overheard
Why did Mark's email mention a redundancy? It was Leanne who lost her job.
The product saved a note about Leanne under your account, not in her file. Delete it in your settings. I'll ask the vendor how we stop one client's details reaching another client's work.
Where it sits
Memory
- is part ofHarness
- works withModelChat app vs AI inside a productWhere client data goesContextTraining
Grey words are still being written.
Copy this entry
Paste it into Word, Google Docs, Notion or Apple Notes. Headings and links come with it.
Sources
Your feedback
Could you explain memory to a colleague now?