AI Memory: How It Works and How to Make It Last
AI memory is whatever an assistant can see about you and your work when it answers. It comes in four layers. The context window covers this chat only. Built-in memory is the facts an app saves about you, and projects hold files and instructions for one topic. External notes are the fourth: notes the AI reads through a connection such as MCP. The first three live inside one app. Only the fourth is long-term memory you own and every AI can share.
If you use one assistant for light tasks, built-in memory is enough. If you use several, or your work depends on details like a brand voice or a client brief, the fourth layer is the one that holds up.
What is AI memory, exactly?
A language model doesn’t remember anything between requests on its own. Each answer is produced from the text in front of it. “Memory” is the set of tricks apps use to put the right text in front of the model: the chat so far, saved facts, project files, or notes it looks up.
So the useful test is what the AI will see when you start a new chat tomorrow, and who decides what that is.
What are the four kinds of AI memory?
| Layer | What it holds | How long it lasts | Who controls it | Shared across AIs? |
|---|---|---|---|---|
| Context window | The current conversation and attached files | Until the chat ends or fills up | The chat | No |
| Built-in memory | Short facts the app saves about you | Until you or the app deletes them | Mostly the app | No |
| Projects | Instructions and files for one topic | As long as the project exists | You | No |
| External notes (MCP) | Any notes you or your AI save | As long as you keep them | You | Yes |
Context window
Everything in the current chat. It’s the most precise layer, because the model sees your exact words, but it ends with the chat, and long chats push early details out. AI context window explained covers why it fills and what to do about it.
Built-in memory
ChatGPT, Claude and Gemini each save facts about you across chats: your job, preferences, ongoing projects. It’s convenient and mostly automatic. The catch is that the app decides what’s worth saving, the entries are short, and each app keeps its own. Claude memory vs ChatGPT memory and Gemini memory vs ChatGPT compare them.
Projects
A project bundles instructions and files for one topic, like a client or a product launch. It’s the best layer inside a single app for keeping contexts apart. Claude Projects vs ChatGPT Projects explains the differences.
External notes
Notes stored outside the assistant, which it searches, reads and writes through a connection. MCP, the Model Context Protocol, is the standard way to do this now: Claude, ChatGPT and Cursor all support MCP connections (what is MCP). Of the four layers, this is the one you control in full, because the notes are plain files: you can read every word, edit it, export it, and connect more than one AI to it.
Why does built-in AI memory feel unreliable?
Three reasons come up most:
- It’s selective. The app picks a few facts and drops the rest. The detail you needed (the exact wording of your offer) is often the one it didn’t keep.
- It fills up. Built-in memory has limits. ChatGPT memory full shows what happens when you hit them.
- It’s stuck in one app. What Claude knows, ChatGPT doesn’t. Switching assistants, or using two, means starting over.
None of this makes built-in memory bad. It’s good at small personal facts. It just isn’t built to hold your working knowledge.
How do you give AI long-term memory?
Keep the things you reuse in notes, and let your assistants read and write them.
- Pick what’s worth remembering. Brand voice, offers, audience, decisions, recurring prompts, an idea bank. Not every chat.
- Write each as a short Markdown note with a clear title, so search finds it. Markdown for beginners takes ten minutes.
- Connect the notes to your AI through an MCP server, so it can search before it answers and save when you say “keep this”.
- Link related notes with
[[wikilinks]], so the AI can follow from a launch plan to the offer it depends on. - End sessions with a save. “Save what we decided as a note” is the habit that makes the memory grow.
For a Claude-specific walk-through with four options compared, see how to give Claude long-term memory. To make it work across assistants, see share memory between Claude and ChatGPT.
What is an AI memory layer?
It’s the shared store that sits beside your assistants and holds what they should remember, instead of each app keeping its own. Some memory layers are built for developers, as a library inside their own app. Others are for people: a hosted vault your existing AIs connect to. If you build software and want memory inside your own product, a developer library is the better pick. If you want Claude and ChatGPT to remember your work, you want the second kind.
FAQ
Do AI assistants remember past conversations?
Some do, partly. ChatGPT and Claude can refer to past chats and keep saved memories, depending on your plan and settings. They remember summaries and facts, not every word.
Is AI memory private?
It depends on the app’s settings and policies. You can usually view and delete saved memories in each app’s settings. Notes in your own vault are visible to you in full, which built-in memory isn’t always.
What is persistent memory for AI?
Memory that survives beyond one chat. Built-in memory and projects are persistent inside one app. External notes are persistent across apps.
Can different AIs share the same memory?
Not through their built-in memory. They can share external notes, if each one is connected to the same notes through MCP.
Where Valta fits
Valta is the external-notes layer: a vault of Markdown notes with an MCP server. Connect it to Claude (claude.ai, the desktop app or Claude Code), ChatGPT, Cursor or VS Code, and each one saves, searches, reads and links the same notes. No AI runs on Valta’s side, so it doesn’t decide what to keep. You and your assistant do. What you get: you explain your offer once, and tomorrow’s chat in a different AI starts from it instead of from zero. On the web, each vault shows up as a graph of linked notes, with a reader and editor.
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | 1 vault, up to 100 notes, graph view |
| Pro | $9 a month or $90 a year | Up to 5 vaults, unlimited notes (fair use 10,000), version history |
| Team | $8 per seat a month, from 3 seats | Pro for every member, plus shared team vaults |
Pro and Team start with a 14-day free trial, then the price above; cancel before it ends and you pay nothing. Prices are low because a Valta call costs milliseconds of server time, not model fees. Export your Markdown on every plan, Free included. Details are on the pricing section.
Start your free vault and give every AI you use the same memory.