The one benchmark miss ("what css framework" → a memory that only says tailwind) is cross-domain synonymy with no corpus evidence — unreachable for a purely lexical system.
Proposal: an optional MIND_EMBED_CMD (or Ollama URL) hook used only for re-ranking, with graceful fallback to the built-in hash embeddings. Hard constraints: mind.py stays single-file stdlib-only; offline behavior unchanged and default; the benchmark gains a second column (offline vs. embedded) so the gap is measured, not claimed.
The one benchmark miss (
"what css framework"→ a memory that only says tailwind) is cross-domain synonymy with no corpus evidence — unreachable for a purely lexical system.Proposal: an optional
MIND_EMBED_CMD(or Ollama URL) hook used only for re-ranking, with graceful fallback to the built-in hash embeddings. Hard constraints:mind.pystays single-file stdlib-only; offline behavior unchanged and default; the benchmark gains a second column (offline vs. embedded) so the gap is measured, not claimed.