Thoughts on memory layers like mem0
Really curious if there is actually a market for companies trying to sell a memory layer for AI agents.
Look at Claude code's "automemory" for example, they don't use a vector database or anything complex. Its just simple markdown files on the local filesystem along with grep for retrieval and it works great for me.
For many simpler use cases, simple techniques like context summarisation are good enough.
8 comments
[ 0.24 ms ] story [ 10.0 ms ] threadEspecially now that building a simple memory layer particular to your use case won't take a lot of time and effort. And you can keep making it better as you go forward.
I do think it's actually a Hard problem, otherwise everyone would have done it. ChatGPT is the king at it, but they're not selling memory as a product, so I assume it's tightly coupled with the chat. Claude can't do it without token guzzling.
And the first thing I set up with AI chatbots is memory. You need memory to work with past conversations, and suddenly you're dealing with emergent behavior and token drain.
The other question is timing.
It's probably a year late. You could easily make a billion dollar company just from this, but just as easily be disrupted by Big AI.
There will probably be a time when this would be unbundled, similar to how telcos and cloud unbundle all kinds of offerings. And that opportunity will come several years later.
You also need to handle duplicate or conflicting facts, and that becomes a burden as the memory grows.
Many app builders might start using a vector database. This will help with storage and organising data, but will fall short on keeping an accurate memory when facts change.
So I think there is a market, but only for the multi-user case.