“It remembers me” can describe several different things. A system might have the current conversation in view, load a saved preference, retrieve an old document, or produce an answer influenced by training. These are different places to look when asking where information is kept and what a deletion control changes.
The following four-part map is a simplified explanatory model. Products combine these functions in different ways, and the map does not describe the storage policy of any particular service.
The conversation in view
Material included in a model's current input can affect the next response. That does not establish that the same material will be included next month. Conversely, disappearing from the current conversation view does not prove that every stored copy has been removed.
Saved records and preferences
An application can store a statement such as a preferred appointment time and supply it again later. That record may be editable. The important question is whether editing it also updates related copies, summaries or instructions. An interface label alone cannot answer that.
Material retrieved when needed
A service may search an archive and bring selected passages into a response. Retrieval-augmented generation describes a technical approach to combining retrieval with generation; the foundational paper distinguishes external retrieved material from learned model information. Lewis and colleagues.
Information reflected in training
Training changes model parameters. Removing a source document from an archive is a different operation from addressing its influence on a model trained with it. Machine unlearning concerns that latter problem. These differences matter even when the screen presents one continuous conversation.
Ask about a particular record
Imagine a service that knows your old address. Correcting a profile might fix future delivery suggestions while leaving an old invoice unchanged. Keeping that invoice may be legitimate and useful; treating it as your current address would be another matter. Ask which record the answer used, how it is dated, and what correction is available.
Sovereign Mercy's Chapter Ten explores retained knowledge after access ends. CAELUM's persistent learning and reasoning architecture belong to the fiction; they should not be assumed to describe a present-day chatbot. The Right to Be Wrong offers a smaller version of the distinction through an account connection and its history.
Continue with Can an AI forget what it has learned? to separate ending access, deleting records and changing learned behavior. The goal is a more precise question, not the reassuring or alarming claim that AI either remembers everything or forgets on command.
Listen with this question in mind
- Griefbots and the Digital Afterlife
If access to the service ends, what would the family still have?
Sources and limits
The examples and recommendations are editorial explanations. Fiction illustrates the questions; it is not evidence of real-world outcomes.
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Foundational architecture; not a description of every modern RAG product or a griefbot study. - Machine Unlearning
Does not demonstrate universal deletion in all architectures; no experimental speed-up numbers reused.