A saved assistant remembers instructions, not every past chat

Start here

A reader who repeats the same kind of task — drafting weekly reports, tutoring on one subject, reviewing the same kind of document — eventually asks whether an assistant can remember the setup rather than being re-explained every time. OpenAI answered this first, with GPTs in November 2023; Anthropic later did the same with Projects. Both are a container for instructions and files, not one continuous memory across every conversation, and reading their own descriptions closely prevents the common mistake of assuming a saved assistant knows what happened in an unrelated chat.

What the documents say

OpenAI's 'Introducing GPTs' announcement, dated 6 November 2023, describes GPTs as 'custom versions of ChatGPT' that combine 'instructions, extra knowledge, and any combination of skills,' and states that 'anyone can easily build their own GPT' with no coding required. OpenAI's help centre page on Projects in ChatGPT, retrieved 16 September 2026, describes a related but separate feature: a project keeps files, instructions and chats together for one piece of work, and lets a user 'add project instructions' that 'only apply inside the respective project' and override the account's global custom instructions while inside it. Anthropic's support article on Projects, retrieved the same day, describes Projects as 'self-contained workspaces with their own chat histories and knowledge bases,' notes that 'enhanced project knowledge with RAG is only available to users with paid Claude plans,' that free accounts are capped at five projects, and states directly that 'Projects maintain separate chat histories and knowledge bases per project' rather than forming 'a unified persistent memory across all Claude conversations.'

Check this

The concrete check is to open two different projects, or GPTs, built from the same account and ask each a question that depends on something discussed only inside the other. If the assistant cannot answer, that confirms the documented boundary: instructions and uploaded files are scoped to the container, not to the account as a whole. A reader can also check whether their plan is free or paid, since Anthropic's own page ties retrieval capacity and project limits directly to the subscription tier.

What holds and what fails

The 'saved setup, not saved history' model holds across both vendors' current documentation and is a genuine capability for repeatable tasks. It fails if a reader expects the saved assistant to also remember new facts from casual conversations elsewhere in the app, or expects free-tier limits to match a paid account's. Both companies' pages describe current settings that can change; this is not a permanent architecture guarantee.

• Decide what one repeat task actually needs — a fixed instruction, or reference files — before building a GPT or project around it.
• Test whether the saved assistant can see outside its own project, rather than assuming it inherits other chats.
• Check a plan's project or GPT limits on the vendor's own help page, since free and paid accounts differ.

A GPT or a Project is closer to a labelled folder with instructions attached than to a second memory for the whole account, and both vendors' pages make that boundary explicit rather than implied.