Enterprise Search
AI Agent Memory for Business: Company Knowledge, Facts and Open Cases Compared
An AI agent without memory starts from scratch every day. But “memory” means something different depending on the vendor: the knowledge in your organisation’s data, facts about people and agreements, or an overview of open cases. I separate these three areas, show the approaches on the market and give you five questions to measure any offer against.
by Ing. Christoph Wendl · · 9 min read
AI-generatedA note up front: we at Iphos IT Solutions in Vienna make searchit and are therefore part of this comparison ourselves. About the other vendors, we only write what they publish themselves, as of November 2026. My advice: check every statement against your own data – with us as with any other vendor.
A language model can formulate and draw conclusions, but it knows nothing about your company. It knows neither the framework agreement with your most important supplier nor what was agreed last week. Nor does it know that the quote to Ms Berger has gone unanswered for six days. For an AI agent in your business to do useful work, it needs a memory. The question is: which one?
In short
- Three memories, three questions. Company knowledge (what is in your data), facts (what applies to people, customers and agreements), open cases (what is outstanding). For each one, permissions, provenance and currency have to be settled.
- The market builds fact memory in two ways. Some capture a lot from conversations and let the user tidy up afterwards. Others store more selectively and check at the moment of creation.
- Team memory is now widespread. What matters is who sees it – and whether the permissions of the sources travel with it.
- Does the memory sit inside one agent or beneath all of them? Anyone using several AI agents doesn’t want to teach each one separately what applies in the organisation.
Three memories that are often confused
| Memory of … | What it contains | Where it comes from | The most important question |
|---|---|---|---|
| company knowledge | contracts, quotes, minutes, emails, records from line-of-business applications | from the systems where it already sits today | Does everyone see only what they are allowed to see in the source system? |
| facts | “Mr Huber is the contact for maintenance”, “Customer X only wants quotes as PDF”, preferences and responsibilities | from conversations with the agent and from completed work, corrected by hand | Where does the fact come from, and is it still valid? |
| open cases | who is waiting for an answer, what was promised, which deadline is running | from emails, appointments and tasks | Does the system recognise when something is done? |
The first row is the classic field of enterprise search and AI-powered knowledge management. If you want to use AI for knowledge management in your company, you’ll find an overview of vendors in our post Best Enterprise Search Software 2026. Here I go one step further: how do facts and open cases come in – and how do all three work together?
How the market builds memory for AI agents
Facts: capture and tidy up afterwards – or store selectively
With the big AI assistants, the agent itself remembers details from conversations. Microsoft Copilot stores them in the user’s mailbox. According to Microsoft, the user can view, edit and delete them in the settings at any time. Administrators, however, cannot restrict what kind of information goes in. With ChatGPT, users also manage their memories themselves. Google describes Gemini Enterprise as a system that learns from every interaction. Memory building blocks for developers such as Mem0, Letta and Zep also analyse conversations in the background.
This has an advantage: the agent quickly feels familiar. At the same time, sooner or later the question arises in a company: what is actually in there, where does it come from, and who confirmed it?
The second approach stores more selectively: only what the user asks for, what applies permanently, or what they confirmed after a suggestion – each with source and date. This way, every fact can be traced back and corrected.
Team memory: widespread, but protected differently
Shared knowledge for teams is now advertised by many vendors, including Notion, Atlassian Rovo, ThoughtSpot, Writer, Zep and Cognee. They differ in how they handle permissions. Coworker.ai, for example, writes that the access rules of the sources travel with every fact. With GitHub Copilot Memory, only users with read access to the repository see the memories. With others, team memory is a separate store with its own permission model.
Open cases: the youngest of the three fields
A system that recognises from emails, conversations and appointments what is outstanding is still new. There are announcements on this from Interloom, Asana and Salesforce, for example, and from Microsoft under the name Autopilot. Some of this is still in a pilot phase or has so far only been announced.
Five questions for any agent memory
- Whose permissions apply? Does an agent see in the memory only what its user is allowed to see – for documents, for facts and for team knowledge?
- Where does an entry come from? Does every fact carry its source, date and the agent or person who created it?
- Who can correct? Can the user view, change, delete or time-limit a fact – and does the previous state remain traceable?
- How long is it valid? Does the memory know validity periods, and can you query what applied on a given date?
- For which agents? Does the memory sit inside a single agent – or do all the AI agents your employees use find it again?
On top of that comes an additional question for IT: who decides what goes in? For facts about customers and employees, it is part of the broader question: how can AI be used in a business in a GDPR-compliant way?
How searchit builds memory
searchit is the memory beneath your AI agents. It sits centrally in searchit, and every connected agent finds it again.
Company knowledge: your data itself
- What is already there: file servers, mailboxes and line-of-business applications from more than 100 source connectors are indexed and enriched with AI; from this, searchit recognises people, companies, places and amounts together with their relationships. An agent queries this knowledge graph directly.
- With the permissions of the sources: with every query, the permissions of the source systems apply – right into the knowledge graph. About documents the user isn’t allowed to see, their agent learns neither names nor relationships.
Facts: central, with source and validity
Since version 8.0, searchit has a central memory for facts:
- How facts are created: the agent writes a fact when you ask it to (“Remember …”) or when you say something lasting – a preference, an agreement, a responsibility. From completed work, searchit suggests facts; they are only created once you confirm.
- What every fact carries: it applies personally, for a role or for a project, and carries its source, date, the writing agent and a validity period. Duplicates are merged, and facts can be queried as of a given date.
- Who sees them: reading follows the same permissions as the rest of your data. Writing requires its own role, and every write is logged with its author.
- You keep track: on the “Memory” page, you see your facts by area. There you can edit, delete or time-limit them. The previous state remains traceable.
- Specialist knowledge per role: specialist knowledge can additionally be assigned to a role as a knowledge package with version and validity.
Open cases and projects: what is outstanding
- Open cases: searchit recognises from your own emails where you are waiting for an answer, what you have promised and which deadline applies. The other party’s reply is recognised too. If searchit recognises a case incorrectly, you correct it with one click; the correction takes effect in the next run.
- Project memory: for a project, you create a profile – with goal, participants, deadlines, rules and decisions. Only members see the project; anyone who leaves loses access without any further step.
One memory for all agents
The searchit AI Agent, Claude Code, OpenCode and Codex get the same tools: query and edit open cases, search and write facts, fetch the context of a project. What an employee has taught one agent, the next one finds again – in each case with the user’s permissions.
What this looks like is shown on the slide “One memory for all agents – everyone sees only their part” on searchit-agent.ai.
Which approach suits you
When looking for an AI tool for knowledge management, it really doesn’t matter whether “memory” is written on the box. What matters is how you want to work.
If you use a single assistant and want it to pick up as much as possible from conversations, you are well served by the big AI assistants. If you have several AI agents in-house, need to keep facts about customers and employees traceable, and want to use the knowledge in your systems with their permissions, I recommend a memory beneath the agents – central, with source, validity and correction.
Frequently asked questions
What is memory for AI agents?
Everything an agent draws on beyond the individual conversation: the knowledge in the company’s data, facts about people and agreements, and the overview of open cases. Without memory, an agent knows only what is in the current conversation.
What distinguishes agent memory from knowledge management?
AI-powered knowledge management makes your data findable. Agent memory goes further: it also records what was agreed in conversation and what is outstanding – and passes both on to every agent working for an employee.
How does a fact memory fit with data protection?
Facts about people are often personal data. Source, purpose, validity and the ability to correct and delete entries are therefore important. This post is not legal advice; check with every vendor who stores what and who may see it.
Can several AI agents use the same memory?
Yes, if the memory sits outside the agents and can be reached via a common interface. At searchit, that is the Model Context Protocol (MCP), an open standard for AI tools.
Conclusion
Memory for AI agents in business covers three areas: company knowledge, facts and open cases. With every offer, I would ask: whose permissions apply, where does an entry come from, who can correct it, and do all your agents find it again?
We’ll be glad to show you what the searchit memory looks like on your own data, in a demo.
Related posts: Best Enterprise Search Software 2026 · How to Safely Test AI Agents with Write Access
Information on other vendors: as of their own publications in November 2026. Microsoft Copilot and Autopilot are trademarks of the Microsoft group of companies, ChatGPT and Codex of OpenAI, Gemini of Google, Claude Code of Anthropic, GitHub Copilot of GitHub; all other product and company names mentioned are trademarks of their respective owners. Mentioning them does not imply a partnership with the owners or an endorsement by them.
You might also like
AI-generatedEnterprise Search
AI Document Search: Checking PDF Tables in Your RAG System Without the Overall Score Hiding Errors
A RAG system with a very good overall score can still assign the decisive amount wrongly. Especially with multi-page PDF tables, headers, columns and page references get lost …
AI-generatedEnterprise Search
Permission to Read Is Not Permission to Act: How to Safely Test AI Agents with Write Access
An AI agent with read permissions finds information. With write permissions, it sends emails, changes records and creates appointments. Before you allow that, you need three …
AI-generatedEnterprise Search
Acceptance Testing AI Agents for Business: Measure Errors, Clarifying Questions and Rework Separately
In a pilot, an AI agent often looks good. Whether it saves more work in day-to-day operation than review and maintenance cost is something a success rate alone won’t tell you. …
