A hard lesson from building OpsRadar: a vector database is not an agent memory layer.

We see so many founders dumping their company docs into Pinecone, hooking it up to Claude, and calling it an autonomous agent. When it actually runs, the agent pulls conflicting chunks of text, gets confused by old pricing models, and asks the user 10 clarifying questions.

We had to completely rebuild how memory is structured. At OpsRadar, we designed a unified "Company Brain." It gives agents a deterministic rule layer (how to execute) alongside a dynamic history layer (what happened before), completely independent of the LLM provider.

Curious for the builders here: are you still relying on raw RAG for your agent's context, or have you started building dedicated memory states? We are currently optimizing our API response times for heavy state retrievals and would love to compare notes.