Good that it's open source, though always sus about what they actually do. I've seen a lot of supposed memory layers end up just being a RAG system with remember and recall over pg_vector or something. I'll admit I haven't really looked into the code to verify whether that's actually the case so open to being fact checked.
I assure you it is not a vibe coded pg_vector wrapper. It is built on top of a custom made search engine that is state-of-the-art in semantic search. Check out moorcheh.ai, cloud infrastructure, moorcheh on-prem and moorcheh-on-edge.
Here is the paper explaining how we came up with the idea of memanto: https://arxiv.org/abs/2604.22085
Thanks for the reply! So you've actually built your own custom vector DB on top of the memory layer? That's definitely something new.
I'm honestly not sold on memory layers as a whole, I find that adding more context only serves to make the LLM's dumber and I've been slowly leaning towards just having no memory and working on features in isolation.
You've done benchmarks already so that's another step in the right direction, do you find that these benchmarks are good indications of practicality? Have you seen a major difference in LLM performance utilizing this as a memory layer?
To answer your first question, I have to say no. agentic memory benchmarks do not reflect the nuances and difficulties that real agents encounter in real workflows. We hope there are better more practical benchmarks being developed.
Second question I assume by LLM you mean AI agents; if that is the case, yes, the whole agentic work experience is different when you have an active memory agent to rely on.
I first tried it last week, since then i migrated to this seek thing, really good on recall and easy to manage and maintain. We definitely need our data to be owned by us!
We have built our own infrastructure on moorcheh.ai, it was the first information-theoretic search engine before Google's research. Moorcheh whitepaper: https://arxiv.org/pdf/2601.11557
Thanks for your comment but Memanto is not a "memory layer" it is an active "memory agent". The difference is that memanto is not a passive layer it is an actual agent and it has its own intelligence and AI capabilities independent of your agents.
I'm honestly not sold on memory layers as a whole, I find that adding more context only serves to make the LLM's dumber and I've been slowly leaning towards just having no memory and working on features in isolation.
You've done benchmarks already so that's another step in the right direction, do you find that these benchmarks are good indications of practicality? Have you seen a major difference in LLM performance utilizing this as a memory layer?