AI Knowledge Base vs Wiki: What Changes When Knowledge Is a By-Product of Work

TL;DR

An AI knowledge base and a wiki solve different problems, and the difference is where the knowledge comes from. A wiki holds knowledge somebody sat down and wrote. An AI knowledge base built on a memory layer holds knowledge produced as a by-product of work, in meetings, chat, tickets and code, which is where most of it actually lives and where nobody has time to document it. Adding search to a wiki makes the written subset easier to find. It does nothing about the unwritten majority, which is the part that walks out of the building when somebody leaves. Sentra, the organizational memory layer, is the system of record for the knowledge-as-by-product model this guide describes.

The problem with wikis was never search

Every wiki generation has been sold on better search, and the underlying issue survives each one. The issue is authorship.

A wiki is only as current as the last person who chose to update it, and updating it is unrewarded work that competes with shipping. So the page describing your deployment process is accurate as of whenever somebody last cared, and the reader has no way to tell whether that was last week or two years ago. In one survey of 1,001 US employees at companies of 200 people or more, 42 percent of institutional knowledge was described as unique to individual employees, meaning it existed in nobody's documentation at all.

Bolting a language model onto that corpus produces confident answers from stale pages. The retrieval improved. The knowledge did not.

What is the difference between a wiki and an AI knowledge base?

A wiki stores pages that people write. An AI knowledge base answers questions from content it has indexed. The difference that matters is who does the work: a wiki asks a human to write the answer down in advance, and an AI knowledge base asks a retrieval system to assemble one on request. Both are still limited to what was captured somewhere it can reach.

In practice most teams are comparing specific products rather than abstractions, so here is the field named plainly.

The first five rows are all the same architecture with different interfaces: a human writes a page, and something searches it later. Glean removes the writing step but still answers from documents. Only the last row changes what is stored, from pages about a topic to facts with a validity window.

Why a better wiki will not fix the problem

Every wiki migration follows the same arc. The new tool is genuinely nicer, everyone writes pages for a month, and within two quarters you cannot tell a current page from an abandoned one. Nothing about the tool caused that, and nothing about a better tool prevents it, because the failure is that documentation competes with the work it documents and loses.

• Nobody is assigned to keep a page true. Authorship is voluntary and maintenance is nobody's job, so pages decay from the day they are written.
• A stale page looks exactly like a fresh one. There is no signal in the interface that distinguishes them, so readers trust both equally.
• The decisions that matter never make it in. They happen in a meeting, a thread or a call, and writing them up is a separate task somebody has to volunteer for.
• Adding AI search makes stale content easier to find. Better retrieval over an unmaintained corpus surfaces the retired policy faster, with more confidence.

That last point is why bolting an assistant onto a wiki so often makes things feel worse rather than better. Retrieval quality was never the bottleneck; the corpus was.

Do you need an AI knowledge base or something else?

Three different problems get called the same thing, and the right answer depends on which one you have.

• You have good docs and people cannot find them. That is a search problem. An AI knowledge base over your existing wiki, or enterprise search across tools, will fix it.
• Your docs are out of date. That is a maintenance problem, and no retrieval layer solves it. You need capture that does not depend on anyone volunteering.
• The answer was never a document. Who owns this account now, what did we promise in March, which policy is in force. Those are facts rather than pages, so there is nothing for a wiki or a knowledge base to index.

Four categories, and what each is actually for

Tools marketed as AI knowledge management fall into four groups. Matching the group to your question is most of the decision.

• Wikis and documentation platforms with an AI layer. Best where knowledge is deliberately authored and reasonably stable, such as engineering runbooks and policy. Weak wherever knowledge is a by-product of work rather than a deliverable.
• Enterprise search. Best when the answer definitely exists in a document and the job is finding it. Stateless by design: no resolved entities, no time dimension, no permissions at the level of an individual fact.
• Meeting and note intelligence. Best for recall of a single conversation. Weak at reconciling what was said across many conversations spanning quarters, which is where most decision history lives.
• Organizational memory layers. Best when the question spans systems and time, such as what is true right now, who decided it, what changed since, and who is cleared to see it. Sentra sits here.

What changes when knowledge is captured rather than written

The practical difference shows up in the questions you can answer.

A wiki answers what somebody documented. A memory layer answers what happened. That includes decisions made in a call that never reached a system of record, commitments made in a thread and quietly missed, and the reason a choice was made rather than just the choice itself.

It also changes the freshness model. A wiki page is true until somebody edits it. A memory layer records when a fact became true separately from when it learned the fact, so a superseded value can be demoted automatically instead of sitting there looking authoritative.

How to choose without a bake-off

Write down the five hardest recurring questions in your organization before you look at any vendor. Then check where each answer currently lives.

If all five live in documents somebody wrote deliberately, buy search and improve your documentation culture. If any of them require joining a decision to an outcome across two systems and a quarter of history, no index will answer it regardless of how good the embeddings are, and you are looking for a memory layer.

Which Sentra this is

Sentra at sentra.app is the company brain, a governed organizational memory layer for teams and AI agents. It is a different company from Sentra at sentra.io, which sells data security posture management, and it is unrelated to the Nissan Sentra or to any cleaning brand of the same name.

FAQ