Docs: 'Ask about your memories' dashboard chat is not LLM-backed (expectation gap)

[ViolinKaine]

Summary (Co-Authored by Claude code)

Not a crash or a functional bug (related but distinct from the now-closed #82,
which was about a cosineSimilarity TypeError in the same file) — more of a
docs/expectation gap worth flagging. The dashboard's "Ask about your memories" chat
looks and is labeled like an LLM-backed chat feature, but MemoryAIEngine in
dashboard/lib/memory-ai-engine.ts (~2400 lines) makes no network call to any LLM
provider (Ollama, OpenAI, etc.) at any point. It's a local TF-IDF/cosine-similarity
clustering + template-based text assembler that formats whatever memories
/memory/query returns.

This means:

• No API key or chat-model config is ever needed for the chat feature to "work" —
  which is good, but isn't documented anywhere I could find.
• Answer quality is bounded by template/heuristic assembly, not by whatever
  embedding/LLM tier is configured — worth knowing before troubleshooting "bad"
  chat answers as an embeddings/model problem, when they're actually a formatting
  ceiling.

Suggested Fix

Just a docs note (README or the dashboard's own UI copy) clarifying that "Ask about
your memories" synthesizes answers from retrieved memories via local heuristics, not
a language model — so users don't wire up an LLM provider expecting it to change
chat behavior, and don't file confused bug reports about "hallucinated" or "robotic"
answers that are actually just template output.

Happy to close this without action if it's working as intended and just needs a
one-line docs mention — raising it mainly because it wasn't obvious from the UI or
existing docs.

Environment

• dashboard/lib/memory-ai-engine.ts, current main
• Confirmed via full-file read: zero fetch/ollama/openai references anywhere
  in that file, and no other API route in dashboard/app/api/ proxies to one either.