Long-term memory in LLMs is usually a joke—either it's a tiny context window that forgets your project requirements after ten prompts, or it's
RAGthat pulls irrelevant chunks of data. Hermes Agent claims to actually "remember" workflows and previous solutions, which sounds great on paper, but I'm skeptical about how it handles state management without hallucinating old versions of a codebase.If this is just a fancy vector database wrapper, it's not a revolution. However, if it can actually maintain a persistent, evolving understanding of a specific AI workflow across different sessions, it solves a massive pain point for anyone doing complex prompt engineering.
I'm trying to figure out if this actually reduces the "context priming" phase we all go through every time we start a new chat. Does it truly integrate a persistent memory layer, or is it just augmenting the prompt with a few retrieved snippets? For those who have done a deep dive into the architecture, is this actually a reliable LLM agent for production, or is it just more marketing fluff?
Next NumPy: My first real encounter with array computing →
All Replies （3） #
J
Usually depends on how they handle the metadata tagging for retrieval, otherwise it's just noise.
0
D
My last RAG setup just hallucinated random docs, so I'm curious if this actually sticks.
0
N
Does it use a vector database for the long-term memory or some custom caching?
0