save-memorieslisted

## Save Memories to cc-memory Database Review the current conversation and persist **structured memories** through the **anti-patch writer** (`llm.memory_writer.upsert_smart`). Never bypass this path: it auto-decides whether to MERGE (overwrite an existing high-similarity memory), SUPERSEDE (archive an old version and insert a refined one with a chain link), or INSERT (genuinely new fact). This prevents stacked duplicates. ### Step 1 — Review the conversation Identify, since the last save or session start: - **Decisions** made (architecture choices, parameter selections, tradeoffs) - **Results** (measurements, benchmarks with specific numbers + units) - **Bugs** found and fixed (root cause + fix; especially "NEVER do X" warnings) - **Config** changes (hyperparameters, env vars, settings that were tuned) - **Architecture** insights (module structure, pipeline design, data flow) - **Tasks** still pending or blocked ### Step 2 — Score each candidate - `category`: one of `decision`, `result`, `config`, `bug`, `task`, `arch`, `note` - `importance`: 1-5 (5=critical/never-forget, 4=important, 3=useful, 2=minor, 1=skip) - `content`: one self-contained sentence with **specific values** (numbers, file paths, parameter names). Bad: "tuned the learning rate". Good: "lr=3e-4 picked over 1e-3 because val_loss flatlined after epoch 8." - `topic`: a short lowercase keyword for grouping (e.g. `auth`, `pipeline`, `ui`) ### Step 3 — Quality bar - Only **conclusions**, not the discuss