Why memory matters more than the model
Right now, most teams pick an AI vendor based on which model is smartest this month. That choice changes constantly. New models arrive every few weeks, and the leader board flips. If your strategy depends only on which model is on top, you have to redo your work every time the field moves.
Memory is different. Memory is the record of what your customers want, what your team has tried, and what worked. Memory grows over time. The longer you run, the richer it becomes. A great model with no memory is a stranger every morning. A modest model with deep memory feels like a colleague who has been with you for years.
Three layers of useful memory
Useful memory has three layers. The first is short term, which is what the agent remembers inside a single task. The second is account level, which is what it remembers about you and your team. The third is ecosystem level, which is what it learns across many users while still keeping each person’s data private.
Most products today have only the first layer. They handle one chat well and forget everything else. The second layer is appearing in calendars, email clients, and code editors. The third layer is rare. Building it requires careful design, strong privacy controls, and a platform that touches many parts of a user’s work life. That is the prize worth chasing.
The portability test
Here is a simple test you can run on any agentic product. Ask whether your memory can move with you. If you leave the product, can you take your history, your preferences, and your learned routines somewhere else? If the answer is no, you are renting your own memory. If the answer is yes, the product is treating memory as a shared asset.
We think portable memory will become a standard expectation, much like portable phone numbers. Regulators in Europe and Asia are already discussing data portability rules for AI services. The companies that get ahead of this shift will be trusted. The companies that fight it will look like the carriers who once tried to lock customers into one phone book.
Current events shaping memory
Several current events are pushing memory to the center of the conversation. New copyright cases are forcing companies to log exactly which sources their agents read. New privacy rulings are forcing them to let users delete history on demand. New enterprise contracts are requiring audit trails that show what the agent saw, decided, and did.
All of these forces point the same way. Memory cannot be a hidden side effect. It must be a first-class product feature, with its own controls, its own export tools, and its own service level. The teams treating memory as an afterthought will be caught flat-footed when the next round of rules lands.
How to design for memory now
If you are building anything agentic, design for memory from day one. Decide what to remember. Decide what to forget. Decide where it lives. Make those choices visible to the user. Treat the memory store as carefully as you treat the database that holds payments.
Done well, memory becomes the part of your product that competitors cannot copy. They can match your features in a sprint. They can match your model with a contract. They cannot match years of careful, consented memory built with your customers. That is the moat that survives every new model launch.