What is a knowledge base in AI?

Short answer

A knowledge base in AI is the store of information a system reasons or answers from. In classic AI, such as the expert systems of the 1970s and 1980s, it held facts and rules that an inference engine applied. In today's AI tools it usually means a set of documents, notes or records that a language model searches before it answers, a method called retrieval-augmented generation (RAG).

Two meanings of the same word

The term is older than ChatGPT, and both meanings are still in use. Which one someone means usually depends on whether they studied AI before or after 2020.

Classic AI (expert systems) Modern AI (LLM tools)
What is stored Facts and if-then rules in a formal language Documents, notes, help articles, records
Who reasons over it An inference engine applying the rules A language model reading the retrieved passages
How it is built Knowledge engineers write the rules by hand You upload or connect existing text
Strength Exact, explainable conclusions Works with ordinary language and messy documents
Weakness Brittle outside the rules someone wrote Can pick the wrong passage or an outdated version

The classic meaning: knowledge base plus inference engine

In computer science, a knowledge base was originally one of the two halves of an expert system. The knowledge base held facts about a domain and the rules that connect them. The other half, the inference engine, applied those rules to the facts to derive new ones. It could work forward from what is known (forward chaining) or backward from a conclusion it tries to prove, asking the user for any fact it is missing (backward chaining). Expert systems appeared in the 1970s and were widely seen as the future of AI in the 1980s. Large knowledge bases from this tradition, such as Cyc, and later ones like Wikidata and ConceptNet, still exist.

What made them a knowledge base rather than a database is that the content is meant to be reasoned over, not just looked up. Wikipedia also notes a second, softer use of the word: knowledge management products call any repository for human readers, like manuals and policies, a knowledge base.

The modern meaning: what an LLM retrieves from

A language model knows what it absorbed in training, and nothing about your company. The 2020 paper that named retrieval-augmented generation (Lewis et al.) described it as two kinds of memory: the model’s own parameters, and a separate index of documents that a retriever searches. That separate index is what people now call the knowledge base.

In practice, it shows up in every major tool under a different name:

  1. Claude Projects: “project knowledge”, the files you upload into a project. On paid plans Claude switches to RAG automatically when the files approach the context limit.
  2. ChatGPT: the knowledge files of a custom GPT, or the files in a project.
  3. Microsoft Copilot Studio: “knowledge sources”, such as SharePoint sites, uploaded documents, Dataverse, public websites and connectors.
  4. Support tools such as Zendesk and Intercom: the help center articles the AI agent answers customers from.

The mechanics differ per tool. LLM tools can place files directly in context or retrieve selected passages; Claude Projects automatically enables RAG when project knowledge approaches context limits. Either way, the model writes the answer from what it was given, ideally with citations.

The misunderstanding: putting it in the knowledge base means the AI knows it

Uploading a document does not teach the model anything. The model sees only the material the application places in its context. That may be a whole file or retrieved passages. If three versions of a price list sit in the knowledge base, it may pick the old one. If a rule is buried on page forty, a question phrased differently may never reach it. And nothing in the knowledge base changes the model itself: delete the file and the knowledge is gone.

That is why the quality of an AI knowledge base depends less on its size than on its curation: one leading version per topic, clear titles, and someone who removes what is outdated. More on how to set one up in How do you build a knowledge base for AI agents? and on the difference with memory in Is RAG the same as AI memory?

What this means for a team

Most teams end up with several small knowledge bases without deciding to: a project in Claude, a custom GPT in ChatGPT, an agent in Copilot Studio, each with its own copy of the handbook. Teams can connect several tools to shared sources, such as SharePoint, but still need to manage source freshness and decide how a correction made in one tool becomes shared knowledge. A knowledge base also holds documents, not the things a team learned along the way: which answer was wrong last week, which customer gets a different rate, which way of working was agreed. Those live in people’s heads or in chat threads, and the AI never sees them. For the document side, see How do you make company documents searchable for AI?

Where Bombos fits

Bombos is one shared memory underneath all the AI tools a team already uses, instead of a separate copy per tool. It is a plugin for Claude, Claude Code, ChatGPT, Copilot, Grok, Codex, Cursor and your own agents that learns how your team works, draws on what already sits in mail, documents and repos, and shows which sources an answer rests on. It runs in the cloud in Europe or standalone on your own server, and one person can try it first with their own agents.

Related questions

Is a knowledge base the same as a database?

No. A database stores records for software to look up. A knowledge base stores what is known in a form a system can reason or answer from: rules and facts in classic AI, readable documents and notes in LLM tools.

Is a knowledge base the same as training data?

No. Training data shapes the model once, before you use it. A knowledge base sits outside the model and can be updated any day; the model looks things up in it at the moment you ask.

What is the inference engine in a knowledge-based system?

The part that applies the rules to the facts to derive new conclusions, either forward from known facts or backward from a goal it tries to prove. In LLM tools the language model itself takes over that reasoning role.

Does a team need a separate knowledge base for every AI tool?

Most tools ask for one: a Claude project, a custom GPT, a Copilot Studio agent each hold their own sources. That works for one tool, but it means the same knowledge is copied and kept up to date in several places.

Sources

  1. Knowledge base (Wikipedia)
  2. Expert system (Wikipedia)
  3. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., arXiv 2005.11401)
  4. What are projects? (Claude Help Center)
  5. Knowledge sources summary (Microsoft Copilot Studio, Microsoft Learn)
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.