How to choose a company memory tool: the eight criteria that decide
Disclosure: we make a tool in this category. So read the criteria below as what they are, an argument about what memory means, made by someone with a stake in the answer. Every one of them is written to be tested by you, in a trial, rather than believed.
The reason this page is not a list of vendors: the category moved faster than any list can. Wikis added agents, search tools added citations, developer libraries added temporal graphs, and new funded entrants appear every quarter. A ranking published today is stale in a month, and every one of these products now uses the word "memory" on its homepage.
What does not go stale is the set of questions. Here are eight, in the order they matter.
The question underneath all the others
Before the criteria, the one that sorts the field in a single sentence:
What does this tool do when nobody asks it anything?
Storage waits. Search waits. An agent fires on a rule somebody wrote in advance. A memory is the only category that owes you an answer you did not request. Hold that question in mind through everything below.
The eight criteria
1. Does it date the validity of a decision, or only its edits?
Why it decides. A version history tells you the text changed. It does not tell you whether the decision still holds. Six months later, that gap is the whole problem: the page looks current, and the decision behind it is dead.
What a good answer looks like. Every decision carries a start, what it replaced, and the evidence it rested on. You can ask "what did we believe in March" and get March's answer, not today's.
How to test it. Record a decision. Change it. Then ask the tool what the earlier state was and why it changed. If all you get is a diff, this is version control, not memory.
2. Is every claim tied to the evidence it rests on?
Why it decides. A memory that cannot show its work is a confident stranger. And a decision detached from what motivated it can only be re-argued, never re-evaluated.
What a good answer looks like. Every answer cites its sources, and abstains when it has none rather than filling the gap.
How to test it. Ask a question the tool cannot know the answer to. A good one says so. A weak one produces a fluent paragraph.
3. Does it surface anything nobody configured?
Why it decides. This is the criterion most tools fail, and the one that costs the most. What hurts a company is never what it thought to watch: the objection returning for the twelfth time with nobody counting, the March decision the field has contradicted since June. Nobody writes an alert for that, because you do not know in advance it is happening.
What a good answer looks like. The tool tells you something you did not ask for, and it is right.
How to test it. Feed it two weeks of real material, then do not ask it anything. See what it says on its own. Many trials never do this, which is why many trials end in a tool nobody opens.
4. Can it link two things that share no vocabulary?
Why it decides. Search returns what resembles your question. But cause and symptom almost never use the same words: the churn reason and the pricing decision that caused it look nothing alike on the page.
How to test it. Put in a customer complaint and, separately, a decision that plausibly caused it, with no shared terms. See whether anything connects them.
5. Does time exist in it?
Why it decides. "This objection has doubled since last month" is not a search result, it is a measurement. A tool built on pages cannot produce it, because a page is a snapshot.
How to test it. Ask what is rising and what is fading. If the answer is a list of recent documents, time does not exist in this tool.
6. What does it cost to keep it fed?
Why it decides. Every system in this category works in month one. The ones that survive are the ones that do not require somebody to file, tag and link by hand. Structure holds as long as someone pays that cost, and it collapses on the first busy quarter.
How to test it. Count the actions between "this happened" and "the tool knows it". If it is more than one, multiply by a year.
7. Can you leave with it?
Why it decides. A memory is the least portable thing a company can lock up, and the most expensive to rebuild. This applies to AI assistants above all: what a model learns about your company should not be trapped in a subscription you might cancel.
What a good answer looks like. Full export, and a protocol like MCP that lets several AI tools read and enrich the same memory without owning it. Open source and self-hosting are the most direct version of this guarantee, when your situation calls for them.
How to test it. Export on day one of the trial, not day ninety. Look at what you actually get.
8. Where does it live, and who can compel access?
Why it decides. Hosting is not a technical detail once your memory holds your decisions and your customer conversations. Data sovereignty is about which jurisdiction can compel disclosure, not about where the marketing page says the company is based.
How to test it. Ask for the hosting region in writing. A clear answer is a good sign in itself.
What not to choose on
Three criteria that look decisive and are not.
- The number of integrations. Two hundred connectors matter only if you use the six that hold your actual knowledge. Check those six.
- The size of the template gallery. It measures the community, not the memory.
- Whether the homepage says "memory". By 2026 they all do. That is precisely why this page exists.
Three profiles, three honest answers
You mainly need to write and document together. Take a knowledge base. Criteria 1 through 5 will not matter to you, and no memory product will replace simultaneous editing and documentation governance. The difference between the two categories is set out separately.
You are building a product whose core is memory. Take a developer building block and keep control of the engine. Some of them do the temporal model very well. You will be writing the ingestion, the interface and the retrieval yourself, which is several months, and that is the right investment if memory is your product. The detailed comparison says which does what.
You are a founder or a small team, nobody is writing an integration this quarter, and the same debates keep coming back. Then criterion 3 is the one that decides, and it is worth testing before anything else.
Using this page against us
The criteria above are the ones Verbasil is built on, so of course we score well on our own grid. The useful thing to do with that is to turn it around: run these eight tests on us too. Criterion 3 is the one we would want to be judged on. Criteria 7 and 8 are the ones we answer with no caveats. On documentation governance and cross-tool search we do not compete, and we say so on our comparison pages.
Book a memory audit: free, 30 minutes, we look together at what your company has already forgotten.
FAQ
How do you choose a company memory tool?
Test eight things rather than comparing feature lists: whether it dates the validity of decisions, ties claims to evidence, surfaces things nobody configured, links items that share no vocabulary, measures change over time, costs little to keep fed, lets you export and connect other AI tools, and hosts your data somewhere you accept. The third is the one most tools fail.
What is the difference between a company memory tool and a wiki?
A wiki answers "where is the document". A memory answers "what should come back to me now". The wiki is an organised writing space you consult; the memory dates facts, links things that look nothing alike, and comes back to you unprompted. Most teams need both.
How do you test a memory tool during a trial?
Feed it two weeks of real material, then deliberately ask it nothing for a few days and see what it tells you on its own. Most trials consist of asking questions, which only tests search. Also export your data on day one rather than at the end, so you find out early what leaving actually looks like.
Do you need an open source memory tool?
Only if you need to read the code, modify it, or host it yourself, which are real requirements in some regulated contexts. Otherwise what matters is portability: a full export and a protocol that lets several AI tools read the same memory, so that no single subscription holds it.
Which criterion matters most?
Whether the tool surfaces anything when nobody asks it anything. Storage, search and rule-based agents all wait for you to act. What costs a company dearly is what nobody thought to watch, and that is the one thing only a memory can return.