When an AI "remembers", it's doing very different things depending on what we're looking at. Worth keeping the levels distinct, because important choices depend on which memory we're actually using.
I identify four.
1. Parametric memory
The model's general knowledge, learned at training time. Lives in the weights of the neural network, millions or billions of numerical parameters tuned on large text corpora.
What it remembers: general facts about the world, language rules, reasoning patterns, domain knowledge up to the training cutoff.
Where it lives: inside the model itself, distributed across parameters.
Who controls it: whoever trained the model.
How long it lasts: forever, but doesn't update without retraining. It's frozen.
Key limit: it knows nothing about you specifically. It doesn't know what you did yesterday.
2. Context memory
The prompt window: what the model sees in this specific call. Includes the user's message, recent conversational history, any attached documents, any content retrieved from an external system.
What it remembers: everything inside the window, nothing outside it.
Where it lives: temporarily in memory during the request.
Who controls it: whoever composes the prompt — the app, the developer, partly the user.
How long it lasts: the time of a single call. Call ends, it disappears.
Key limit: as huge as it's become in modern models (hundreds of thousands of tokens), it's ephemeral. And every new prompt rebuilds the window from scratch.
3. Session memory
The memory that some AI platforms (ChatGPT, Claude, Gemini, etc.) offer as an integrated feature: a collection of facts about you that the system saves and re-injects into later conversations. It's the feature that makes ChatGPT remember you have a dog named Lola, or that you prefer replies in English without having to repeat it every time.
What it remembers: usually preferences, specific facts, style examples. The precise structure varies a lot from platform to platform.
Where it lives: in the chat provider's systems, tied to your account.
Who controls it: mainly the provider. You can see and delete entries, but the format and selection logic aren't yours.
How long it lasts: as long as the account exists, and as long as the provider doesn't change the rules.
Key limit: it's closed in the platform. Switch provider, lose everything. Switch model inside the same provider, often it's not carried over. It's app memory, not personal memory.
4. External persistent memory
A vault of your content, living outside the AI, that AIs can consult via protocol (in Timo: MCP).
What it remembers: everything you write into it — notes, documents, fragments, decisions, project contexts. Structured however you want.
Where it lives: wherever you decide — on your device, on your own server, on a dedicated service.
Who controls it: you. You export, modify, inspect, delete without asking permission.
How long it lasts: as long as you maintain it. Independent of the AI of the day.
Distinctive feature: it's cross-AI. The same memory serves one model today and a different model tomorrow, with no migration.
A map, not a replacement
The four levels don't compete, and they don't replace each other. They work together.
Parametric memory gives the model general knowledge of the world: without it, it couldn't even speak your language. Context memory is the present moment, where processing happens. Session memory holds small conveniences between conversations in the same app. External persistent memory is the long-term storage of your knowledge, reusable outside the single app.
When you hear "this AI has memory", it matters to ask: which of the four levels? The answer changes a lot.
Cheat sheet: which level, for what
| You need... | The level that matters |
|---|---|
| To write better | 1. Parametric memory (change model) |
| Continue a long line of reasoning in the same session | 2. Context memory (wide windows) |
| Remember small preferences across chats | 3. Session memory |
| Not rebuild context on projects that last months or years, across different AIs | 4. External persistent memory |
Practical consequences
The first step in choosing AI tools sensibly is understanding what you really need from them. Often "I want an AI with memory" means very different things depending on the person. For someone, level 3 (session memory) is enough. For those working with continuity on long-running projects, level 4 is needed.
It's not the same choice. And it can't be solved by the same tool.
Try Timo free for 15 days — timoai.xyz. No credit card required.