Knowledge Graph or Vector Store for Company AI Agent Memory
Knowledge graph vs vector store for AI agent memory: which records suit each, how to keep them in sync, and how to test that memory stays current.

Use a vector store for company text that must be found by meaning, such as manuals, policies, call notes and long procedures. Use a knowledge graph for records that have identities and relationships, such as customers, jobs, assets, approvals and the owner of each fact. Most company agents need both: the graph decides which entity a question is about, and the vector store finds the passage that answers it.
Key takeaways
- Vector stores find text by similarity of meaning. They are the right home for documents that are read as passages.
- Knowledge graphs store entities and the relationships between them. They are the right home for facts that must be linked, dated, owned and corrected.
- Memory compounds only when every write has a source, a version, an owner and a way to be expired or corrected.
- Store the passage in the vector layer and the fact pointing to it in the graph, rather than copying the same text into both.
- Judge the design with retrieval tests built from real staff questions, not by the name of the database.
Why the storage choice shapes what your agent remembers
When a business owner hears "vector database," it is easy to assume the agent now has a memory. In practice, a vector store holds numerical representations of chunks of text and returns the chunks that look most similar to a question. That is useful, but it does not know that the "Mercer account" mentioned in an email and the account code on an invoice belong to the same customer. It also does not know that a warranty rule was replaced last month, unless someone wrote that down in a way the system can use.
Drift usually looks like this: the agent retrieves a plausible passage that was true in an older version of a policy, or it attaches a fact to the wrong customer because two records look alike. Neither failure is dramatic, and both erode trust slowly. The storage choice matters because it determines which of these failures you can detect and which you can fix.
What a vector store does well
A vector store is strong when the question is about language rather than identity. A staff member asks how to handle a damaged pallet at delivery, and the answer lives in a three-page procedure that nobody has summarized. Chunking the procedure, embedding the chunks and retrieving the closest ones gives the agent a good chance of finding the right paragraph even when the staff member's wording differs from the document.
The weaknesses are predictable. Chunks can split a rule from its exception. Similar passages from different document versions can compete with each other. Nothing in the store says which passage is current, who approved it, or which customer it applies to. For these reasons, vector retrieval works best on documents that are reviewed, versioned and written to stand alone. The practical details of splitting long documents are covered in Chunk Size for AI Knowledge Bases: Keep Policies and Procedures Whole, and the general question of whether a vector database is needed at all is covered in Do You Need a Vector Database for Company AI Agents?.
What a knowledge graph does well
A knowledge graph represents things and the links between them. Nodes might be a customer, a job, an asset, a location or a staff role. Edges say that a job belongs to a customer, that an asset is installed at a location, or that an approval was given by a manager. Each edge can carry a date, a source and a status, which makes the graph much easier to audit than a pile of text.
The graph answers questions that similarity search answers poorly. Which open jobs belong to this customer and its parent company? Which approvals are still valid for this asset? Which fact replaced the previous one? Traversal and exact lookup handle these questions directly. The cost is upfront design: you must decide the entity types, the identifiers and the rules for matching records. Extraction from documents can also introduce errors, so graph facts need the same review as any other memory. A practical starting scope is described in Company Knowledge Graphs for AI Agents: Start With Five Entity Types.
A decision guide: which records go where
Use the following questions to sort each kind of record. If the answer to the first question is yes, the record probably belongs in the graph as an entity or fact. If the answer to the second is yes, it probably belongs in the vector layer as a passage.
- Does the record describe a specific thing the business acts on, such as a customer, job, asset, site or contract? Put it in the graph with a stable identifier.
- Does the record need to be linked to other records, such as an approval that applies to one job only? Put the link in the graph.
- Does the record need an expiry date, an owner or a correction history? Put the fact in the graph, even if its source text lives elsewhere.
- Is the record a long explanation, procedure, policy or set of examples that is read as language? Put the passage in the vector store, with source, version and date attached.
- Is the record a casual observation from a conversation that has not been reviewed? Keep it out of both until it is checked, following the review rules in Checking AI Agent Conversation Summaries Before They Become Memory.
Illustrative scenario: a regional equipment rental company receives a question about a delivery fee for a crane booked last spring under a different company name. The vector store retrieves the rental terms passage. The graph shows that the booking belongs to the same customer entity, linked to a parent company, and that the terms were replaced in a later version. The agent answers with the current version and cites it, and staff can see why the older passage was set aside. Without the graph, the agent would have quoted terms that looked right but were no longer in force.
A step-by-step plan that keeps memory from drifting
The design works only if it is operated. The following sequence is a sound starting point for a first engagement, and it can be run by a small team with clear ownership.
- Step 1: List the entities the business acts on every day and give each one a single stable identifier, even if the source systems use different codes.
- Step 2: Assign an owner to each record type, and write one rule for who can create, change and retire it.
- Step 3: Store every document passage with its source, version, effective date and review date, so the agent can say which version it used.
- Step 4: Link graph facts to the passages that support them, rather than copying policy text into the graph.
- Step 5: Collect twenty to thirty real questions from the people who answer customers, and record the passage or fact that a correct answer must use.
- Step 6: Run those questions before and after every change to memory, and log each miss with its cause, such as a missing link, an outdated passage or a duplicate entity.
- Step 7: Review expiring facts on a fixed schedule, and remove or correct anything that no longer matches the business.
Several of these steps depend on clean source documents. If the documents are scattered or unversioned, start with How to prepare company documents for an AI agent before building either layer.
How Be Human runs this memory design
Be Human designs and operates personalized AI infrastructure in which an agent keeps its identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries, and an observable success criterion. For memory, that usually means deciding which records belong in a graph, which passages belong in a vector layer, and who owns each write. Managed operations then cover monitoring, recurring retrieval evaluations, integration maintenance, permission review, incident response and controlled improvements, so the stores are checked against real questions instead of being left to drift.
For a business owner, the next decision is usually narrower than choosing a database. It is choosing one role, one set of records and one question the agent must answer correctly every time. Start there, measure it, and expand only when the first role behaves as expected. Before you choose, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Is a vector store enough for company AI agent memory? For unstructured documents that people read as text, a vector store is often a good first layer. It struggles with identity, relationships, dates and corrections, so most businesses eventually need a structured layer for those records.
When is a knowledge graph worth the extra effort? A graph becomes worth the effort when the agent must connect customers, jobs, assets or approvals, or when it must know which fact replaced another. Start with a small set of entity types and expand only after retrieval tests show a real gap.
Can one system do both jobs? Some platforms offer both capabilities in one product, and that can simplify operations. The design question remains the same: the system needs passages that can be found by meaning, facts that can be linked and expired, and a clear owner for every write.
How do I know whether my agent's memory is improving? Measure it with a fixed set of real questions, the correct source for each, and a log of misses. Improvement means fewer wrong answers and fewer stale citations over time, not more stored content.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


