Using AI for Book Recommendations: ChatGPT, Gemini and What Memory Actually Does
ChatGPT and Gemini are far more capable than Highstreett at almost everything. Both can recommend books well. Both also have memory that they themselves describe as interpretive rather than literal, which makes them unreliable at one narrow task: keeping an exhaustive list of what you have already read. Neither filters retailers by ownership.
What they are better at
Open ended requests, nuance, tone, follow up questions, and everything that is not this. If you want one assistant for the whole of your life, that is a real advantage and we do not have it.
On shopping, OpenAI added Instant Checkout and the Agentic Commerce Protocol in September 2025 and a shopping research tool in November 2025. In March 2026 it stepped back from standalone checkout and refocused on discovery: visual browsing, side by side comparison, image based lookalike search. OpenAI states shopping results are organic and drawn from publicly available retail sites. These features are live for US users and not confirmed in the UK.
Gemini takes a different route, embedding commerce across Google’s Shopping Graph rather than building a separate shopping mode.
Last checked: 4 August 2026
Memory, described by the people who built it
ChatGPT has two memory systems. Saved memories are explicit and editable. Reference chat history is implicit, and OpenAI’s own guidance is to use saved memories for anything you want it to always keep in mind, because what it carries forward from chat history can shift over time. OpenAI has noted saved memories tend to go stale.
Gemini has saved info plus a cross chat personal context that builds a compressed profile, which Google describes as interpretive rather than literal, and which can be dropped when context limits are reached.
Neither is a database. Both are probabilistic summaries of conversations, which is the right design for a general assistant and the wrong one for a list of four hundred books.
A shelf is a list of rows. It does not drift, compress or go stale. That is not intelligence, it is just the correct data structure for the job.
The retailer question
Neither assistant filters retailers by ownership. Both are moving toward merchant fed catalogues through commercial protocols, which means the retailers who appear are the retailers who build an integration. Independent bookshops are not building agentic commerce integrations, and will not be.
That is a structural reason the independent answer will keep being missing, rather than a temporary gap.
Questions
Does ChatGPT remember what books you have read?
Partly, and not reliably. ChatGPT has saved memories, which are explicit, and reference chat history, which is implicit and which OpenAI says can shift over time. It is not a structured list, so it can recommend something you have already told it you read.
Can ChatGPT buy things for you in the UK?
Not confirmed. OpenAI's shopping features, including Instant Checkout and shopping research, launched for US users. UK availability is not confirmed as of 4 August 2026.
Do AI assistants recommend independent retailers?
Not specifically. Neither ChatGPT nor Gemini filters by ownership, and both are moving toward merchant fed catalogues, which favours retailers large enough to build an integration.
Why does a shelf work better than memory for this?
Because it is a list rather than a summary. Memory in a general assistant is compressed and probabilistic by design. A shelf is rows in a table: four hundred books stay four hundred books.