Why your AI forgets everything between conversations
It is not a bug, and it is not going to be fixed by a bigger context window.
The short answer: a language model has no state between requests. Everything it appears to “know” during a conversation is text sent to it in that request. When the conversation ends, nothing persists, because there was never anywhere for it to persist to.
What actually happens
Each time you send a message, the entire conversation so far is transmitted to the model. The model reads it, produces a reply, and retains nothing. The sense of continuity within a chat is the transcript being re-sent, not the model remembering.
Start a new conversation and the transcript is empty. There is no mechanism by which last week's exchange could be present.
Why bigger context windows do not solve it
A larger window lets one conversation run longer. It does nothing across conversations, because the window is filled from the request — and something has to decide what goes in it.
Longer contexts also degrade retrieval: the fact you need competes with everything else loaded alongside it. A million-token window that you fill with noise is worse than a small one filled deliberately.
The four approaches
- Paste it every time. Works, costs tokens on every session, and you will forget things.
- A project instructions file. Better — a fixed preamble loaded automatically. Fails as it grows, because everything is loaded for every question regardless of relevance.
- Built-in assistant memory. Convenient, and now standard across the major assistants. The limits: it lives in one vendor's product, you have limited control over what is kept, and it does not follow you to a different tool.
- An external memory the assistant can query. Facts stored outside the conversation, retrieved when relevant. More setup, and the only option that survives switching assistants.
Choosing between them
If you use one assistant for casual work, built-in memory is enough and free. Take it.
External memory earns its complexity when you use more than one assistant, when you want the data on your own machine, or when what you need remembered is specific enough that you care what gets stored — corrections, decisions, constraints. That is the case Varven is built for: Markdown on your disk, readable by any client that speaks the Model Context Protocol.
Either way, what you store matters more than where. What belongs in memory covers it.
Common questions
Does Claude remember previous conversations?
Claude has a memory feature that carries context across chats, available on all plans. It is separate from the model itself, which remains stateless — memory works by storing information outside the conversation and supplying it back on later requests.
Can I make ChatGPT and Claude share the same memory?
Not through their built-in features, which are per-vendor. An external memory exposed over the Model Context Protocol can be read by both, since the protocol is supported across assistants.
Is a bigger context window the same as memory?
No. Context is working space for one conversation; memory is what survives between them. A larger window does not decide what to put in it, and filling it indiscriminately makes retrieval worse.