Widemem adds confidence checks for memory

The real bug in most RAG stacks is not recall, it's forced certainty: if the retriever always returns something, the model will happily turn noise into a confident answer. Widemem's confidence tiers are a small change with a big product impact because they let memory become a policy layer, not just a search index.

• –High / moderate / low / none routing gives the app a clean abstain path instead of pretending every hit is trustworthy.
• –A "creative" fallback is useful only when it's explicitly labeled as speculation; otherwise it becomes a hallucination feature with better UX.
• –User corrections are gold: repeated "I already told you this" feedback is exactly the signal you need to pin or re-rank memories.
• –Local-first storage keeps adoption easy, but the lasting differentiation is retrieval governance, not SQLite or FAISS.