# Archived design notes — unwired

Documents in this directory describe design proposals that were never wired
into the runtime. They are kept for historical context only.

| File | Status | Superseded by |
|---|---|---|
| `context_window_attention_spec.md` | Step 4 ("scale QK dot products by item weight") is physically impossible against API-served LLMs (Anthropic / OpenAI / Ollama). The rest was unwired. | `OA_WEIGHTED_MEMORY_PLAN.md` (repo root) |
| `context_window_attention_weights.json` | No loader in any package. Pure documentation. | `.omnius/config/memory-weighting.yaml` (per the new plan) |
| `context_window_attention_model.py` | Python module never imported by the Node.js runtime. | `packages/orchestrator/src/memory/composite-scorer.ts` (per the new plan) |

The current canonical design is `OA_WEIGHTED_MEMORY_PLAN.md` at the repo root,
which replaces the theater with real KV-cache-affecting wiring (Anthropic
`cache_control` markers, Ollama `memoryPrefix` prefix-cache exploitation) and
score-driven materialization tiers.

If you came here looking for "how attention weighting works in omnius," the
answer was: it didn't, until the new plan. Read that instead.