Minerva turns memory into graph search

This is closer to a real memory subsystem than yet another long-chat summary hack. By splitting reads and writes, Minerva gives the model a place to store facts and a separate path to fetch them when context matters.

• –Embedding every node and edge is a smart way to keep the graph searchable, but it only pays off if entity resolution stays clean.
• –The write path is the interesting part: the model emits `FactEdge` triplets directly and a background orchestrator resolves them, which is cleaner than post-hoc extraction.
• –Graph expansion after semantic retrieval is the right hybrid for personal memory: vector recall finds candidates, graph hops recover structure, and together they avoid flat-RAG amnesia.
• –The local stack is the biggest practical win here; running on llama.cpp and a small model set makes private memory plausible without sending your life story to a cloud API.
• –The repo is still experimental and says it is not production-ready, so the hard part now is evaluation, latency, and whether memory writes stay accurate over long sessions. [Reddit announcement](https://www.reddit.com/r/LocalLLaMA/comments/1s1t36i/local_assistant_with_tool_based_memory_knowledge/), [GitHub repo](https://github.com/RinMar/Minerva)