Three rebuilds later: what stuck

The current stack is not the first one. A few things I learned the hard way:

• Memory layers can poison themselves. An early mem0/Qdrant setup collected contaminated entries and schema problems; ripping it out was faster than fixing it.
• Simpler front ends win. Moving from Ollama and Open WebUI to LM Studio and AnythingLLM cut the number of moving parts and the number of things that broke on update.
• Check the runtime before blaming the model. Crashes I’d pinned on the model disappeared when LM Studio switched from Vulkan to a newer CUDA runtime.
• One GPU, one job at a time. An autonomous agent fighting the chat front end for the same card meant neither worked well. That agent is parked until it has hardware of its own.

Every rebuild now starts from a script and ends with a snapshot, so the next one is a command instead of a weekend.