- Services
- Agency
- About us
Platform
Mustang
HR
Live
The Workforce
Procurement
Live
The Controller
Coming Soon
In Build
Growth & Customer Experience
Marketing
How to Build AI Memory You Can Keep When Models Change: Part 1
Share:
How to Build AI Memory You Can Keep When Models Change: Part 1
Publish date
Share:
Publish date
Share:
What your AI is learning, why the documents are the easy part, and why memory should belong to the company.
Most teams have had the same annoying AI problem for months: the system is smart, but every new conversation starts too blank. People paste the same background again, re-explain how a customer is handled, and remind the system which approval comes next. The model may be brilliant, but the company keeps having to introduce itself.
We have lived versions of this in client AI work for years. Getting a model useful was one problem. Getting the company’s reality to survive across tasks, tools, teams, and model changes was another.
That is why we wrote The Amnesiac Genius earlier this year: the first enterprise AI problem was forgetting. The industry is now getting much better at persistent memory. That is good news. It also creates the next question.
The answer matters before the memory looks valuable. A few months of accumulated learning can be annoying to rebuild. A few years can become part of how the operation works.
AI Memory Is More Than Chat History
The word memory is doing too much work right now. In practice, 3 different things get mixed together, and they need different treatment.
Take a very ordinary pricing decision. The standard discount is company truth. A one time exception approved for a strategic customer is learned memory if the AI is allowed to carry that decision forward. The quote being reviewed today is working context.
If those 3 things live in one undifferentiated pile, the system can easily treat yesterday’s exception like today’s policy. That is where memory stops being convenient and starts affecting the operation.
The Documents Are Usually the Easy Part to Move
Imagine the company changes models next year.
The policy document still exists. The CRM still has the account record. Contracts and transaction history still sit in systems the company owns.
The harder question is what happened between those systems and the AI during a year of work.
Did people correct the AI? Did managers approve exceptions? Did the system learn that one route works better than another? Did an old decision get superseded?
If your AI setup is retaining those things, that accumulated layer can become more valuable than the original prompt. It can also become much harder to move.
Current research is already moving beyond simple retrieval. Recent work is testing how long running AI systems decide what to add, update, delete, consolidate, or keep separate from belief. An active Internet Draft goes further and proposes provider independent, versioned memory objects with provenance and lifecycle state. The details are technical. The business signal is simple: memory is becoming infrastructure.
Where AI Memory Becomes Vendor Lock In
Vendor lock in usually sounds like a Procurement problem. This version starts much earlier.
A company can move a PDF. It can export CRM data. Rebuilding 12 months of corrections, approved exceptions, previous decisions, and the context around those decisions is different.
The risk is not that switching models becomes impossible. The risk is that every month of useful AI work quietly increases the amount of company learning that has to be reconstructed if that learning lives inside one product.
The Rule We Build Around
We have a simple framework for this in our own client work and architecture.
It sounds obvious once written down. It is much harder to retrofit after memory has already spread across assistants, chat histories, local databases, and one off workflows.
The point is not to make every piece of memory portable in every imaginable way. The point is to make sure the company is still the owner of the intelligence that makes the AI useful.
Why We Built Mustang This Way
We kept hitting the same pain in real AI work: models change fast, while company rules, customer history, pricing logic, and operating judgment cannot keep starting over.
That is one reason we built Mustang, our AI operating system. It sits above the model layer so the company layer can stay stable even when the model underneath changes.
Inside Mustang, Grip is simply the knowledge foundation. It holds the company’s live knowledge: rules, customer specifics, pricing, operating knowledge, and senior judgment. Kernel is the part that reads the task, loads only the knowledge that work needs, and checks the result before it reaches a person or a system.
The names are ours. The principle is the important part: the company intelligence should stay with the company, while GPT, Claude, Gemini, or the next model remains replaceable.
Part 2: How to Keep AI Memory Useful Without Letting It Become a Trap
Part 1 is the ownership problem. Part 2 is the operating problem.
What should become memory at all? What should expire? How do you keep a correction separate from source truth? What needs a source and an approver? And how do you know whether your architecture can survive a model change before 2 years of memory are sitting on top of it?
In Part 2 we turn the framework into 5 practical rules and a small model switch test you can run before scale.
Frequently Asked Questions
AI memory is the information an AI system keeps from earlier work so it can use it again later, such as corrections, preferences and previous decisions. It builds up through daily use and often lives inside the tool that collected it.
It depends on the tool and its settings. Many AI assistants now keep some details between conversations, such as preferences, corrections and background on your work, while others start fresh each time. Where memory is switched on, it is usually stored inside that one product and stays there if you change tools.
Most AI tools only see the current conversation plus whatever is placed in front of them. Once a chat ends or the context window fills, earlier details drop away unless a memory feature or connected knowledge source brings them back. Repeating the same background every time is a common sign that memory is missing or switched off.
Often only in part. Documents and CRM records move easily, but what the AI learned through use, such as corrections, approved exceptions and past decisions, usually sits inside the original tool and may have no export option. Keeping that knowledge in a layer your company controls makes switching models far simpler.
It can be, with controls in place. Decide what is worth remembering, record where each item came from and who approved it, and give outdated items a way to expire. Without those controls, a one-off exception can start behaving like company policy. Check each vendor's terms on storage, access and deletion as well.
Related OP reading
- The Amnesiac Genius: Why Your Smartest AI Forgets You Every Morning
- Your Guide on How to Prepare Your Company’s Data for AI
- Your Guide on How to Prepare Your Company’s Systems for AI
Sources
- Association for Computational Linguistics, Memory R1, ACL 2026.
- Association for Computational Linguistics, Hindsight: Structured Agent Memory that Retains, Recalls, and Reflects, ACL 2026.
- Internet Draft, Architecture and Data Model for Persistent Memory in Agentic Systems, July 2026. Early technical proposal, not an adopted standard.
- Optimum Partners, Mustang Overview, 2026.
Share:
Related Insights
Optimum Partners Launches Local AI Innovation Councils to Tackle the Hard Questions of AI Implementation
We spent years trying to stop AI from forgetting the company. Now memory is improving fast, and a new problem appears: where does that accumulated company learning live, and can you keep it when the model changes?
Future of Cybersecurity Teams: AI + Human Expertise for Scalable Defense
Cybersecurity is evolving rapidly. Threats are becoming more sophisticated, attacks happen in real time, and modern IT environments are massive. Traditional approaches alone can’t keep up. The future belongs to hybrid defense systems—teams that combine AI’s speed and scale with human judgment and expertise.
Your Old SaaS Is Now Charging You for AI Add-Ons. The Math Is Worse Than You Think.
Enterprise SaaS renewals are quietly becoming AI consumption contracts. The more value your agents produce, the more you pay vendors to rent the intelligence layer built on top of your own business logic.
Your Guide on How to Prepare Your Company's Systems for AI
We’ve built AI directly into processes running on top of exactly this kind of mess, in both, without tearing out what already works. Ripping out a system that’s run a business for fifteen years, just to
Working on something similar?
We’ve helped teams ship smarter in AI, DevOps, product, and more. Let’s talk.