Connect Pith's hosted MCP server and your assistant answers from your saved sources — with citations — instead of you re-pasting context into every chat.
Your AI assistant is brilliant and amnesiac. Every new chat starts from zero: it forgets what you read last week, and when it does reach for a source it can invent one. Pith fixes the memory side. You bookmark what you read; Pith builds a cited reading memory from it and exposes that memory to your assistant over a hosted MCP server. Now "what do I actually know about this?" gets answered from your own sources, with links you can open — not from a model guessing.
What changes for you
Scenario 1
Stop re-pasting the same context into every chat
Instead of copying three articles into the prompt so Claude understands the topic, you just ask. With Pith connected over MCP, your assistant pulls the relevant passages from what you've already saved and answers from them — the context is already there, cited.
Scenario 2
Ask your assistant what you've read, and get sources back
"What did I save about the new EU reporting rules?" Your assistant queries your Pith reading memory and answers from the actual bookmarks, each claim linked to the page it came from. You get your own reading back, not a plausible-sounding summary of the open web.
Scenario 3
Catch the assistant before it invents a source
When the answer is grounded in your saved bookmarks, every claim carries a link you can open and check. If something isn't in your reading memory, that's visible too — so you're working from what you actually read, not a confident fabrication.
Founder's note
I wanted my assistant to know what I'd read, not start every conversation as a stranger. The model is the thinking; the reading I've done is the memory it should think with. Pith is that memory — saved by me, cited, and handed to whatever assistant I'm using, so the answers come from my sources instead of thin air.
FAQ
How do I connect Pith to my AI assistant?
Pith ships a hosted MCP server — a standard endpoint your assistant can call. In a client that supports MCP (Claude desktop, claude.ai, Cursor, and a growing list), you add Pith's MCP server URL and authorize it once. From then on the assistant can search and read your Pith reading memory as a tool, with no copy-pasting on your side.
Does Pith replace my AI assistant?
No. Pith is the memory, not the model. Your assistant still does the thinking, drafting, and reasoning — Pith just gives it a persistent, cited recall of what you've read so its answers are grounded in your sources instead of starting from nothing each time.
What does "with citations" actually mean here?
Pith's reading memory is built citation-first: each thing it surfaces links back to the bookmark it came from. When your assistant answers from Pith, those source links come through too, so you can open the original and verify a claim before you trust it.
Can this stop my assistant from making up sources?
It changes where the answer comes from. Grounded in your saved bookmarks, every claim has a real link you can check, and gaps are visible rather than papered over with invention. Pith can't police what a model says when it's off on its own, but for questions about your reading it gives the assistant real sources to answer from.
Where is my data stored?
In Frankfurt, Germany — EU-only residency, isolated per workspace, with full export anytime as Markdown or JSON. We do not train models on your content. Your reading memory stays yours and stays in the EU, even though the assistant querying it may run elsewhere.
Do I need to be technical to set this up?
No. Connecting an MCP server in a supported assistant is adding a URL and approving access once — closer to linking an app than writing code. After that, you save bookmarks normally and ask your assistant questions normally; the memory connection works in the background.
Related
Last reviewed: 7 June 2026