Company Knowledge for AI Agents: One Source, Every Agent

Company knowledge for AI agents is a searchable body of decisions, procedures and facts that lets agents answer from what your company actually knows. Your company already has a brain. It is just not written down. The job is not to buy a chatbot for each tool. The job is to write that knowledge down in one place and let every agent look it up the same way. By the end you know the four parts and have your first entries written. Steps 1 and 2 take an afternoon.

What company knowledge for AI agents is

A knowledge base locked inside one chat tool is the next silo. The pattern has four parts, none tied to a vendor:

This guide is part of Loops We Run. The pillar, how we run our company with AI agents, shows the rest.

Knowledge lives in one place, a pipeline keeps the index current, and every agent reaches it through one MCP server.

You do not train a model

"Train the AI on our data" is the wrong picture. A good agent looks things up before it answers, like a new hire with the binder open. That lookup is retrieval: the agent sends the question to a search tool, reads the passages that come back, and answers from them with the source. Nothing is retrained. To change an answer, change the entry.

Meet Larchmoor Cycles, the made-up shop in our examples

Every example uses Larchmoor Cycles, a fictional bike shop with a repair workshop and rentals.

Step 1: Pick one job and list the questions it must answer

Start from the questions people already ask. A knowledge base built from a document inventory answers questions nobody has. Pick one job and write down this month's questions about it from customers and staff, in their words.

Larchmoor picks rentals and returns. Maren's list:

1. Can I return a helmet I tried on in the shop?
2. How much is the deposit on a rental bike?
3. What happens if I bring a rental back damaged?
4. Can I get a refund instead of store credit?
5. How long does a brake job take?
6. Are you open on Sunday?

Keep the list: it is the test in Step 6.

Step 2: Put what your company knows in one place

Write it down the moment it happens, one fact per entry, or it never gets written. Today what Larchmoor knows lives in Maren's head, her inbox and three old documents.

One place holds two kinds of content, and together they are your company knowledge base for AI agents. Documentation covers the procedures, policies and prices you already write: returns-policy, rentals, repair-prices, opening-hours. Captured entries hold the decisions, procedures and facts people learn as they work. Each entry has a title, the fact, why it holds, an owner, a last checked date, a scope, and a valid until date when it expires.

title:  Rental deposit is 100 per bike
fact:   Every rental bike takes a 100 deposit, refunded when it comes back undamaged.
why:    Chose 100 over 50 because a bent wheel costs about 100 to replace.
owner:  Maren
scope:  public

Each entry holds one fact an agent can use alone, with its scope and, where it expires, a date.

For a full data review, read AI data readiness. For memory between coding sessions, see why Claude Code forgets context.

Want us to set this up with your team? Tell us which questions your customers and staff ask every week.

Step 3: Make it searchable by meaning

Customers never use your words, so search by words alone fails them.

We split each document into passages of about 8,000 characters, cut at the end of a sentence, with 200 characters shared between neighbours. 8,000 is big enough to hold a whole rule with its exception. The 200 overlap keeps a sentence from being cut in half. Each passage is stored with its source, title and last-updated date, and turned into an embedding.

An embedding is a list of numbers that captures what a passage means. Two texts that mean the same thing get similar numbers, even with no word in common.

Search scores meaning and words together and merges them into one ranking. This is semantic search with word matching beside it. Meaning finds "Protective gear returns" for a customer who typed "helmet". Words still win for part numbers, names and prices. If the meaning search is unavailable, search falls back to word matches and flags the result as degraded, so the agent still answers. Use a search service you already have, or any vector or hybrid index.

The question says helmet and the passage says protective gear; search by meaning still ranks it first.

Step 4: Keep the index current with a pipeline

An index that is a week old answers with last week's prices. The pipeline matters more than the search engine. A long document is re-read on update: when it is first added, or when a person saves a new version and asks the system to re-read it. Nothing re-reads it on its own.

The pipeline runs on two triggers:

Zero passages written is a failure, never "done". The run reports a cause and tries once more. Each passage id is built from its source plus a fingerprint of its text, so running the pipeline twice never creates copies.

Step 5: Give every agent one way in: an MCP server

Every agent connects through one server, and its access, never its question, decides what it can see. MCP (Model Context Protocol) is an open standard that lets an AI agent call tools on a server. A new agent connects without rebuilding the knowledge for it.

The Larchmoor knowledge server has one read-only tool, search_knowledge, which takes a question and a result limit. We return the top 5 passages by default: enough for the agent to compare sources, few enough to read in full. Each result carries its rank, score, title, source, last-updated date and a snippet.

Each entry has a scope field, and the server filters every search by the connecting agent's access (its key or its sign-in). The shop chat sees public only; Teo's Claude and Maren's ChatGPT see public and internal. No customer can ask their way into internal entries.

Any agent that speaks MCP connects with the server's address and a key or a sign-in. An off-topic question returns nothing, and Maren's instruction tells the chat to say so instead of guessing.

Step 6: Ask real questions, one run traced

Test with real questions and fix the knowledge, never the answer. One question from Maren's list, traced:

1. A customer on larchmoor.example.com asks the chat: "Can I return a helmet I tried on in the shop?"
2. The chat calls the MCP server. Its key carries scope public.

search_knowledge(query="return a helmet tried on in the shop", limit=5)

3. The server returns 3 ranked passages:

The agent asks one MCP tool and gets ranked passages, each with its source, before it writes the answer.

4. The chat answers from passage 1: "Unused protective gear can be returned within 30 days with the receipt. The policy does not cover gear tried on in the shop, so I have passed your question to the team." Source: Returns policy.
5. That is a gap, not a bad answer. Maren decides and writes one entry, scope public: "Helmets tried on in the shop can be returned within 30 days if unworn outside." Why: customers try sizes in store.
6. The entry is indexed on save. The index goes from 40 to 41 passages.
7. The same question now ranks Maren's entry first, and the chat answers from it with its source.
8. Teo asks his Claude, scope internal: "How do we set up tubeless tyres on a rental bike?" The procedure entry ranks first. Nothing was rebuilt for the second agent.

A question passes when the answer quotes an entry, names its source, and that entry is the only place the rule is written. Run every question on your Step 1 list this way.

What goes wrong, and the rule each failure produced

Every failure we hit was a quiet one: something looked indexed or current and was not.

The review list also takes any entry used in 5 or more searches and not checked for 90 days. Turning failures into rules is its own loop: build a self-improving AI system.

Two entries that disagree give two answers; replace the old one, re-index, and date the facts that expire.

Want us to build this for your company?

The first job is the hard part. After it, writing entries is a habit. At Neomanex we build the knowledge base, the pipeline and the MCP server with you, connected to the agents you already use. Your team keeps writing the entries. Tell us which questions come up every week.