Your Chat Memory Is Not Your Library

Claude can remember that you prefer short answers and which project you are working on. It won't remember the talk on retrieval you watched in August, or the three articles that changed your mind about pricing. That gap is what this post is about, and it matters more once you use more than one assistant.

I build BrainTube, which sits on one side of this gap, so read the comparison as labelled.

What built-in memory is for

Claude's memory is on Free, Pro and Max (free memory since 2026-03-02) and ChatGPT's memory reached free users in June 2026; both remember facts about you inside one vendor, not a library of your saved sources that every AI can query.

That design is good at what it is for. Your assistant learns your role, your tone, the project you keep coming back to, and it stops asking the same setup questions. For a lot of people that is the memory they want, and it costs nothing extra.

Two limits follow from the design. The memory lives inside one vendor, so Claude's memory is invisible to ChatGPT and the other way round. And it holds facts about you, distilled from conversations, rather than the sources you learned from. If you want the detail of how each one works and how to edit it, I wrote that up separately for Claude and for ChatGPT.

What a library is for

A library is the other half: the videos, podcasts, articles and PDFs you decided were worth keeping, stored with their full text so you can search them later.

BrainTube compiles what you watch and read into a knowledge base every AI can use. Tools change. Models cycle. Your mind's index stays.

In practice that means three things. It saves the source, not a remembered gist of it. It keeps that source outside any single assistant. And it lets the assistant you happen to be using today search it.

Who needs both

BrainTube is for people who learn from hours of video, podcasts and articles and then lose it, and for people who use several AI assistants (Claude, ChatGPT, Cursor, Gemini) whose built-in memory stays inside each product.

If you use one assistant and mostly want it to know your preferences, built-in memory may be enough. If you switch between assistants, or your work depends on material you read and watched weeks ago, you need somewhere for the sources to live.

How the two fit together

The pattern is simple. The assistant keeps its own memory of me. The library keeps the material. The assistant reaches the library over MCP when a question needs it.

BrainTube runs a remote MCP server at https://mcp.brain-tube.com/mcp; in Claude.ai you add it as a custom connector and sign in with OAuth, with no API key to paste.

Once connected, the assistant does more than read. Your AI can save what it finds into BrainTube: in the same MCP session it can search your library and also save content, add notes and tag items. A useful answer, or a source the assistant found while researching, goes back into the library instead of disappearing with the chat.

This is tested with Claude. Other clients that speak MCP over streamable HTTP use the same endpoint, but they haven't each been checked by hand yet. ChatGPT support is coming; it isn't available yet, so for ChatGPT today the built-in memory is what you have.

Other ways to get a library your AI can use

Having an MCP server is no longer a differentiator: Recall, Readwise, Glasp, Notion, Fabric, Hjarni, MemoryPlugin, Granola, Tana and Supermemory all ship one (as of 2026-10-04). Which one fits depends on what you save.

Readwise (2026-10-04) has an official hosted read+write MCP server, but no free plan (30-day trial); Full costs $9.99/month yearly or $12.99/month monthly. It is strong if your library is highlights and articles.

Hjarni (2026-10-04) is a Markdown knowledge base that AI assistants can read and write over MCP; Free holds 25 notes, Pro is €11/month; it does not ingest video or podcasts.

BrainTube's angle is media. BrainTube saves YouTube videos, web articles, PDFs, EPUB and MOBI ebooks, podcasts and audio, voice memos, Notion pages, screenshots and social posts (Instagram, LinkedIn, X, TikTok), and transcribes and indexes each one. MCP access and bt_ API keys are included on every BrainTube plan, including Free; none of the 38 tools has a paid gate.

What to check before you trust any of them

Whatever you pick, ask the same questions. Can the assistant write to it, or only read? Can you get the whole library out again? Where is it stored?

For BrainTube the answers are: read and write over MCP; export your library as Markdown or JSON from Settings, or the whole corpus through MCP; and your BrainTube library is stored in the EU (AWS eu-west-1, Ireland, via Supabase). The AI providers that transcribe and embed content process it outside that database, so the accurate summary is stored in the EU and private to your account. BrainTube doesn't train models on your library.

FAQ

Does ChatGPT remember the documents I read?

No. ChatGPT's memory holds facts about you from your conversations. To ask about documents and videos you saved, they need to be in a library the assistant can search.

Can Claude and ChatGPT share the same memory?

Not their built-in memory, which stays inside each product. They can share a library if both connect to it over MCP. BrainTube works with Claude today; ChatGPT support is coming.

Is built-in assistant memory enough?

For preferences and context about you, often yes. For the sources you learned from, you need a library with the full text.

Where does BrainTube store my library?

In the EU, in AWS eu-west-1 (Ireland) via Supabase, private to your account.