Reading memory for your AI: Pith's hosted MCP server

Your AI assistant is brilliant and amnesiac. It can reason about almost anything, but it has no idea what you read last Tuesday. Ask it about the regulatory shift you bookmarked three weeks ago and it will either shrug or, worse, confidently invent a plausible-sounding version.

That's not a model problem. It's a memory problem. The assistant has the whole public internet up to its training cutoff and nothing about your actual reading.

Pith fixes the memory, not the model. You bookmark what you read. Pith auto-builds a cited wiki and per-client briefings from it. And now Pith ships a hosted MCP server so your assistant can query that reading memory directly — with citations.

The assistant is the brain. Pith is the memory.

Here's the framing that matters: Pith is not trying to replace Claude or ChatGPT, and it's not an agent framework. It's the thing the assistant talks to when it needs to know what you've read.

MCP — the Model Context Protocol — is the open standard that makes this clean. It's the port your assistant plugs tools and data sources into. Pith exposes your reading memory as one of those sources. No glue code, no scraping, no copy-pasting article text into a chat window and hoping the context survives.

So the division of labour is simple. The assistant reasons, drafts, argues, summarises. Pith supplies the cited substrate it reasons over. A general assistant guesses about your domain. An assistant with your reading memory cites it.

A model without your reading memory is a very smart stranger. With it, it's a colleague who's read everything you have.

Every answer carries its source

The part consultants care about isn't the chat — it's the trail.

Pith is source-grounded by design. Every claim in your wiki links to the bookmark it came from. When your assistant queries Pith over MCP, those citations come back with the answer. So when the assistant tells you "the deadline moved to Q3," it's not a hallucination you have to fact-check — it's a line you can click straight through to the source you saved.

This is the difference between an assistant you trust with a client deck and one you have to babysit. Ungrounded answers cost you a verification pass on every claim. Grounded answers cost you a glance at the link.

It also changes how you use the assistant. You stop asking "what do you know about X?" and start asking "what have I read about X, and what does it say?" The first question gets you the internet's average opinion. The second gets you your own evidence base, cited.

What this looks like in practice

You're prepping for a client call. Instead of digging through bookmarks and old briefings, you ask your assistant: "Pull everything I've saved on this client's market over the last quarter and give me the three shifts that matter."

The assistant queries Pith. Pith returns the relevant passages from your wiki and briefings, each with its source. The assistant synthesises three shifts — and every one of them links back to the article you actually read. You walk into the call with a memo, not a hunch.

Or you're writing. You draft a paragraph, and the assistant checks it against your reading memory, flags the claim that isn't supported by anything you've saved, and points you at the two sources that are. Drafting and provenance in the same loop.

None of this requires you to leave your assistant. Pith lives behind the MCP connection. You keep your tool; you just gave it a memory.

Built for people whose claims have to hold up

This is for consultants and researchers — people who get paid for being right and can prove it. If you cite a source in a deliverable, you need to re-find it. If a partner challenges a claim, you need the trail. A reading memory your assistant can query, with citations attached, is exactly that trail, made queryable in natural language.

Your data lives in Frankfurt, in the EU, workspace-scoped. The reading memory is yours; the MCP server just makes it reachable by the tools you already use.

If you're already living inside an AI assistant, this is the missing half. The assistant brought the reasoning. Pith brings the memory — cited, EU-hosted, and ready to query. See how it fits together on the page for AI-assistant users, or look at where Pith and a general assistant divide the work in Pith vs ChatGPT.

Give your assistant a memory of everything you've read. Then ask it something only your sources can answer.