Reddit thread maps LLM memory stack

Hot take: the interesting part here is not “which memory tool wins,” but that the best setups seem to be composable pipelines with separate layers for raw capture, structured notes, graph relationships, and retrieval.

• –The post is a useful landscape scan of the current second-brain stack, especially for markdown-native and local-first workflows.
• –It reflects the real bottleneck in LLM memory: persistence is easy to store, harder to operationalize and keep current.
• –The open questions are the right ones: NER vs LLM-assisted extraction, local vs API embeddings, and how to prevent stale or bloated knowledge bases.
• –The repo leans toward practical infrastructure, not just theory, which makes it more useful than another generic “RAG + vector DB” roundup.
• –It is still a community question thread, so the value is in curation and discussion quality rather than a finished product.