AI was not built to learn.

I hear a lot of buzz about AI agents getting “smarter” over time. I certainly understand the appeal. You want the agents to learn from your conversations, and have a sense of “memory”.

AI memory management is actually very challenging, and I want to explain why.

1) AI models are built to be stateless, meaning every input is processed from scratch, and the AI models retain nothing afterwards.

2) To simulate memory, an AI agent needs to add prior knowledge into its context window while you are chatting with it. Context windows are very limited. Frontier models support up to 1 million tokens, with smaller models only supporting 256,000 to 128,000 tokens. This means not everything can be “remembered”.

3) Adding too much memory context to every chat will become expensive from the token usage.

4) An agent has to guess what to remember. Unless you are doing manual memory management, you have to hope what you deem to be important is something the AI agent also deemed to be important.

5) An AI agent is also guessing what is relevant for this current chat. It may use context from information that isn’t relevant.

6) Contradicting information is challenging to handle. A memory management system needs to recognize when two pieces of information contradict. If memory is out of date, it may be accurate if you’re asking about the past, but not accurate today.

7) If an AI agent accidentally adds something incorrect to memory, or creates a bad summary, that will affect the accuracy of its memory.

The next time you ask for your AI agents to “remember”, and get “smarter” over time, you are increasing costs, adding additional overhead for memory management, and risking potentially incorrect information.

Make sure your use-case actually requires memory management. For the vast majority of scenarios I see, a curated knowledge base would serve much better.