Tatara
AI & Agents

Stop pasting context into every chat

The hidden tax of stateless agents, and what changes when every tool reads from the same current brain.

A chat window closing while the page beside it breaks up into dots.

You pay a tax every day that never shows up on an invoice: the few minutes spent re-explaining your company to an AI tool before it can do anything useful. You paste in a paragraph about the customer, a note about the pricing, the objection you keep hearing. You get good work back, and then the window closes and it's gone.

On its own it feels trivial. Across a team and across a year, it adds up to a real cost, and almost nobody measures it.

A longer context window or a better prompt won't fix it. What fixes it is a durable place for the knowledge to live, which every assistant can read on demand. That is what Tatara is: a governed document store your agents connect to over MCP, so the facts get fetched instead of retyped.

Why does every AI chat start from zero?

Every AI chat starts from zero because a chat's context is working memory, not storage. Everything you paste lives inside that one conversation's context window, and when the conversation ends the window goes with it. Nothing about the exchange is written anywhere your next conversation, or your colleague's, can reach.

A blank slate is a feature when you want a neutral assistant, and a liability when you want a colleague who knows your business. The agent forgets because nothing durable connects one conversation to the next, not because it's poorly built.

Pasting context is write-once, read-once. Nothing accrues, so the hundredth paste costs exactly what the first one did.

So the same facts get retyped, slightly differently each time, by different people, into different tools. Small inconsistencies creep in. Nobody can say which version is current, because every version lives in a chat that has already closed.

What does pasting context actually cost?

Pasting context has four separate costs, and only one of them is time.

CostWhat it looks like
TimeMinutes at the start of every session, paid by every person, in every tool, forever.
DriftFive people paste five slightly different versions of the same pricing rule. All five agents answer confidently.
StalenessYou change the refund window on Monday. The paragraph in someone's saved prompt still says the old one in November.
UnverifiabilityAn agent gives a customer-facing answer. Nobody can point at the document it came from, because there wasn't one.

Put your own numbers on the first row. If eight people each spend six minutes a day re-establishing context, that's four hours a week, or about five working weeks a year, spent retyping things the company already knows. Treat that figure as a prompt to measure rather than a finding: the magnitude is a guess, but the shape of it isn't.

The other three rows are the ones that hurt. You can feel lost time. You can't feel drift until a customer quotes your own policy back at you and two people disagree about whether it's current.

Why don't per-tool uploads and custom instructions solve it?

Per-tool uploads don't solve it because an uploaded copy is still a copy, and it stops at the boundary of the tool you uploaded it into. Attach your pricing sheet to one assistant and the agent running in your editor still can't see it. Attach it to both and you now maintain two copies that drift apart at different rates.

Custom instructions have the same shape at a smaller scale. They're durable within one tool and invisible outside it, they hold a paragraph rather than a corpus, and nobody reviews them. The update problem just moves. Instead of pasting the same fact into every conversation, you now update the same fact in every tool.

What you want is one copy, many readers. Anything that produces per-tool copies is a smaller version of the problem you started with.

What is the alternative to pasting?

The alternative to pasting is simple: keep what your company knows in one governed place, and let every conversation read from it instead of being fed it. The agent in your editor and the agent in your chat tool point at the same brain.

The important word there is governed. A shared folder of documents is one copy, but it has no shape a machine can use, so an agent still has to read everything to find anything. In Tatara, every document an agent writes carries a type, a folder, a one-sentence description and one to five tags from a registry. That's what turns storage into retrieval.

How does an assistant read from a brain?

An assistant reads from a brain over MCP, the open protocol AI clients use to call external tools. Tatara runs an MCP server. You connect your client to it with a token, and the brain shows up as a small set of tools the agent can call mid-conversation:

  • search_documents: full-text search across the brain, returning ranked hits with highlighted snippets plus each document's id, path, type and tags.
  • list_documents: browse a folder or the whole brain, newest first or changed since a timestamp, returning lightweight rows with each document's one-line description.
  • get_document: fetch the full text of one document.
  • get_taxonomy: the map, meaning the real folders, the document types in use, and the tag registry. Agents call this once at the start of a session.

Nothing is uploaded and nothing is copied. The agent asks a question, gets the current document and answers from it. If you edit that document ten minutes later, the next question gets the new version, in every tool at once.

Tokens carry a scope. A read-only token can search and fetch and nothing else, which is usually what you want for an assistant that answers questions but shouldn't be editing the source of truth.

What changes on a normal Tuesday

Take one real task: a teammate is drafting a reply to a customer who asked about your refund window.

Before, the sequence goes like this. Open a chat, paste the refund policy from wherever you last saw it, paste two lines about this customer's plan, ask for a draft, get a good one. Tomorrow somebody else does the same thing with a version of the policy they found in a different Slack thread.

After, you just ask for the draft. The assistant calls search_documents for the refund policy and gets the document your team actually maintains. It reads the one-sentence description to confirm it's the right one, fetches it, and writes the reply grounded in it. You didn't paste anything, and neither did your colleague, and both drafts quote the same policy because both read the same document.

Then, on the third day, the policy changes. One person edits one document, and every assistant that reads it is current from the next question onward. There is no fan-out step, because there was never a fan-out.

What still belongs in the chat

Not everything should go in the brain, and pretending otherwise makes the brain worse. The rule is durability.

Durable facts belong in the brain: pricing, positioning, the ICP, the objections, the process you follow when a deal stalls, the reason you made a decision. These are true across conversations and worth someone owning.

Ephemeral instructions belong in the chat: the specific task, the tone you want today, the one-off constraint, the file you're working on right now. Writing those into the brain just fills it with noise that future retrieval has to wade through.

The test: if someone else would need this fact next month, write it down once. If it's true only for the next twenty minutes, type it into the chat and let it go.

Stop paying the paste tax.Give the agents you already use one place to read from.
See how it works →

Get that split right and the cost structure flips. Write something down once and it's available everywhere, forever, to every AI you use. The work compounds instead of evaporating, and you stop being the part of the system that remembers.

THE INVITATION

Give your agents context.

Start free and build the one source of truth every AI you use can read from.

500 documents free / no card required

Angus

Founder of Tatara. Writes about company knowledge, AI agents that do real work, and building shared context a whole team can trust.

KEEP READING

More from the Ledger