The knowledge base you already have

Ask most people to set up a knowledge base for an AI assistant and the first suggestion they get is to write one. Summarise your processes. Put your policies in a doc. Keep it updated.

Nobody does this. Not because they are lazy, but because the knowledge already exists and writing it out again is duplicated work. It is in the invoices, the contracts, the statements, the warranty emails and the supplier threads that have been arriving for years. It is just not anywhere an assistant can reach.

docstore.ai is the shortest path between those two facts. Whatever arrives, we keep as it arrived, read it, and make it searchable through an interface an AI tool already understands. There is nothing to write.

Three doors, all open

The one most people use is email. Every account gets an address. You can CC it, set a forwarding rule, or give it to suppliers directly so their invoices come to it in the first place. The body of the message and every attachment become searchable documents, and the original is kept intact.

The second is upload. Drag a file in and it lands in the same archive as everything else. An agent can do this too, without pushing the bytes through its own context: it asks for a short lived upload link and sends the file straight to it.

The third is a bucket. If files already land in S3, point us at the prefix and the archive catches up. It is idempotent, so running it again does not produce a second copy of anything.

What none of these do is normalise your data into somebody else's shape. An email keeps its email metadata. An uploaded file keeps its filename. We do not flatten everything into a lowest common denominator, because the thing you will want in three years is usually the detail that a tidy-up would have thrown away.

Three verbs, not fifteen endpoints

The part that matters for an assistant is how it reaches all this. We went round this several times and ended up with three tools rather than a catalogue of them.

search filters and full-text searches everything. An empty query is just the feed, newest first. show returns one record with its context: the other documents that arrived with it, and the raw source if the agent wants it. edit writes back, because a summary an agent wrote last month should be improvable rather than frozen.

That is the whole surface. It works over MCP, the open protocol AI tools use to reach real data, so connecting Claude or Claude Code is one command and no integration work. There is no folder structure to design, no tagging scheme to maintain, and nothing to keep in sync.

Your ledger comes along too

An archive of invoices is useful. An archive of invoices sitting next to your accounting system is a different thing.

Connect FreeAgent or Xero, and a session exposes those tools alongside the archive. One assistant, one conversation: it finds the receipt, reads the matching bank transaction, and writes the explanation, flagged for a human to approve. Connect Gmail, and it can read the supplier mail that never got forwarded anywhere.

The same applies to what the assistant learns. It can keep notes in the archive, so the rules you work out together stay put instead of being re-explained at the start of every session.

Start with the boring paperwork

We started with invoices and receipts because they are the documents every business has, in volume, without trying. They are also the ones people most often cannot find.

But that is the wedge, not the ceiling. Contracts, warranties, insurance policies, statements, the planning letter you will suddenly need in three years: they are all documents with a date and a promise on them, and they all belong in the same place.

Give your assistant a memory it can search today and still trust in five years. You already have the contents. It just needs somewhere to live.