Hey! I've been working on agent harnesses and llm related infrastructure for a while now.
I've always felt unsatisfied at how RAG and the various existing memory systems work. One of the core issues I kept running into is that agents often need precise recall, not just semantically similar chunks. And thus, I had an idea to give an llm access to structured memory via SQL, ran some experiments and the results were encouraging. Here are the benchmarks https://ingotdb.dev/benchmarks/ and info on the experiments that were run.
The aim is to provide more accurate memory recall and drop-in document ingestion for people developing agent harnesses/loops for prod systems.
More about the motivations are outlined here https://ingotdb.dev/why/ and are hopefully sufficiently explanatory.
Curious to hear what peoples thoughts are and if the value prop is clear based on the landing pages.
Planning to release a hosted version soon, but I intend to leave it OSS and allow for self-hosted deployments.
Hey! I've been working on agent harnesses and llm related infrastructure for a while now.
I've always felt unsatisfied at how RAG and the various existing memory systems work. One of the core issues I kept running into is that agents often need precise recall, not just semantically similar chunks. And thus, I had an idea to give an llm access to structured memory via SQL, ran some experiments and the results were encouraging. Here are the benchmarks https://ingotdb.dev/benchmarks/ and info on the experiments that were run.
The aim is to provide more accurate memory recall and drop-in document ingestion for people developing agent harnesses/loops for prod systems.
More about the motivations are outlined here https://ingotdb.dev/why/ and are hopefully sufficiently explanatory.
Curious to hear what peoples thoughts are and if the value prop is clear based on the landing pages.
Planning to release a hosted version soon, but I intend to leave it OSS and allow for self-hosted deployments.