Your company brain needs an owner and an expiry date

Two arguments about company knowledge landed on the same day last week. They flatly contradict each other, and I think both sides are half right.

On 10 August, Flo Crivello launched Lindy Teammate and made the case against curation on The Cognitive Revolution. Wikis fail, he argues, because "the moment written documentation is written, it's out of date." So Teammate skips the curation step. Its onboarding routine, which he calls hydration, crawls all of Slack, mess and all, because agents don't mind the mess. Three to five million tokens for a twenty-person company with a few years of history behind it.

Same day, No Jitter published a piece on making institutional knowledge AI-ready that opens with exactly the failure a crawl produces. An agent applies the documented Policy A. The sales team has quietly been running Policy B for months. Kuber Sharma of UiPath puts it about as well as it can be put: "The agent confidently produces a wrong answer based on a wrong source. That's worse than producing no answer. It's an authority failure with an AI signature on it."

Both camps are arguing about volume. How much do you feed the thing? I don't think that is the interesting variable. What decides whether your company brain helps or hurts is provenance: for any sentence the agent reads, does anybody still stand behind it?

Mess makes perfectly good evidence. It makes lethal law. Most businesses are about to build one undifferentiated pile in which those two are the same thing.

Why AI agents give confident wrong answers from internal docs

Retrieval ranks by similarity to the question. It does not rank by standing. A Slack message from March 2024 saying "yeah just waive the fee for anyone over $5k" and an approved rate card from last month look like equally good candidates to an embedding model. The Slack message often wins, because people ask questions in the register of Slack, not in the register of policy documents.

Then the agent does the job it was built for. It answers in a confident voice and it attaches a citation.

That citation is the part that does the damage. A colleague who gets it wrong out loud gets challenged, because everyone can hear the guess in their voice. An agent that produces the wrong answer with a link attached gets believed, and then gets pasted into an email to a customer. The failure isn't that the model lied. It did its job faithfully on a source that had no business being treated as authoritative.

Crawl the Slack anyway

Here is where I part company with the tidy camp, because Crivello is right about the thing they keep glossing over. Wikis really do rot. The curation-first pitch has been made to small businesses for fifteen years and I have never once walked into a twenty-person company with a maintained wiki. If the advice has a nil success rate across that many attempts, the problem is the advice.

The stronger point in favour of crawling is one I don't think either side has made loudly enough. The mess is the only place the Policy B problem is visible at all. A tidy, curated-only brain will answer the Policy A question flawlessly and will never know that sales stopped doing it that way in March. The informal record is a detection instrument. It tells you what your business actually does, which is frequently not what your documents say it does.

So crawling earns its place for a better reason than cost. It buys you something curation cannot buy at any price: a live picture of what your business is actually doing. Issuing rulings is a separate job.

What should go into a knowledge base for AI agents

Two tiers, and they never blur into each other.

The evidence layer is wide and cheap and makes no promises. Slack, email threads, meeting transcripts, superseded drafts. The agent may quote any of it with attribution attached and a date on the front, and may never convert it into a rule.

The canon layer is narrow. Every document in it carries four things: a named owner who is a person rather than a department, a last-reviewed date, an explicit statement of what the document does not cover, and a retirement path.

That third one gets skipped by nearly everybody and it is the cheapest insurance in the system. A refund policy written for online orders will be applied cheerfully to a trade account, because nothing in the document ever said it wasn't about trade accounts. Humans infer scope from context they absorbed in the room where the policy was argued out. The agent was not in the room. So write the boundary into the document: this covers standard online orders, it does not cover rush jobs or trade accounts, and if you are being asked about those, no approved policy exists yet.

Which gives you a rule simple enough to actually enforce. The agent may only state a rule that comes from canon. Where canon is silent, the correct answer is "there's no approved policy on this, though here's what people have said about it in #sales". Nobody ever gets fired over that sentence.

Company brain vs enterprise search for a small business

Enterprise search answers the question "what can be found". A brain has to answer "which version is valid". Those are different products, and the market is starting to say so out loud.

Stack Internal, formerly Stack Overflow for Teams, now sells itself as "the trusted knowledge layer" that captures, curates, validates and delivers knowledge with "your own version of our iconic green check mark", under the banner "where human and agentic curation meet". Meanwhile Tencent Cloud open-sourced a team memory hub whose write-up concludes that governance, not retrieval, is the actual differentiator over standard RAG. New memories there default to private, so sharing is a deliberate act rather than a side effect.

An American developer community and a Chinese cloud provider have landed in the same place from opposite directions. Both treat the standing of a document as a first-class field, on the same footing as its text.

Who owns the knowledge an AI agent relies on

Sharma's checklist in the No Jitter piece is five questions long: who owns it, who updates it, who approves changes, who retires obsolete documents, and who audits the agents when official docs change. For a twenty-person business, answering those is a forty-minute exercise, not a project.

List everything your agents can currently read. Put a human name beside each row. Most rows will come back blank. That blankness is the result, and you didn't have to buy anything to get it.

Which is my objection to the reflex I keep seeing, where a governance problem gets answered with a purchase order. A memory product is not a knowledge policy. It's the same trap as buying tooling before you have mapped the work: the purchase feels like the decision, so the decision never gets made. Context engineering is hardening into a named hiring category now, with budgets and job titles attached, which is broadly good news. It also guarantees somebody will try to sell you the job title before you have answered the five questions.

Keeping an AI knowledge base current without a full-time librarian

Freshness is the one axis nobody has solved yet, including the organisations Redis rates most mature and the vendors selling governance as a product.

Redis published its State of Context Engineering survey on 3 August. It found that 97% of respondents believe context matters and 4% have built for it. Another 39% rate themselves ineffective at scoping memory for agents, meaning they can't say what should persist, for how long, or for whom. And 21% admit they frequently or usually run critical decisions on stale or batch-synchronised data, a figure that climbs to 69% among the organisations Redis classes as most mature. Maturity has not yet bought anybody freshness.

The finding I keep chewing on is a pair: 60% rate themselves ineffective at governing access to AI context sources, while 58% are confident their context systems are governed effectively. Redis calls it confidence detaching from capability, which is polite. The same people are grading themselves as failing and passing on the same question.

Crivello concedes the point himself, incidentally, while describing his own memory agent's cache: "every cache at some point got a TTL. It grows stale." That's the crawl-everything camp acknowledging the exact worry the curation camp is built around.

So stop trying to eliminate staleness and make it visible instead. Two habits do most of the work. Keep canon small enough that one person can reread the entire thing in an afternoon each quarter, which in practice means around 15 documents covering the handful of things your business genuinely makes rulings on. If your approved set runs to 400 documents you don't have canon, you have a wiki wearing a better name.

And make expiry demote rather than delete. When a document blows its review date, it drops out of canon into the evidence layer, where the agent can still quote it with a date attached and can no longer state it as policy. Demotion keeps the audit trail intact while stripping the document of its authority to give orders.

What I would write down first

Before you buy a memory product, before you argue about crawl depth, make one page with three columns: the thing your agents can read, the person who owns it, the date it was last checked. Add a fourth column if you're feeling brave, for what would kill it.

We run our own brain at Azgard and I've done this exercise on it. It was not a comfortable afternoon. But it cost nothing, it took less time than a vendor demo, and it told me which questions our agents were not qualified to answer, which is the only thing on that page that would have shown up in a customer email otherwise.

FAQ

How do I stop an AI agent from using outdated company documents?

What should go into a knowledge base for AI agents?

Do I need a full-time person to keep an AI knowledge base current?