diff --git a/ROADMAP.md b/ROADMAP.md
index 3a8b147..3f0671b 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -1,7 +1,7 @@
 # Roadmap
 
 ## Baseline (Not Roadmap Work)
-- **LongMemEval_s: 64.6%** published with full methodology and per-question evidence in [docs/benchmarks/longmemeval/](docs/benchmarks/longmemeval/README.md) (Zep reports 63.8%; Mem0 ~49%).
+- **LongMemEval_s: 64.6%** published with full methodology and per-question evidence in [docs/benchmarks/longmemeval/](docs/benchmarks/longmemeval/README.md). We don't quote other tools' numbers — published runs differ in response models, pipelines, and dataset revisions; our per-question receipts are committed so the run can be judged directly.
 - `v0.9.0`: released. Complete Memory Operations Protocol (all ten verbs, including true `redact`), Context API v2 (per-section budgets, packing strategies, depth tiers), full export/import round-trip, temporal graph memory + entity resolution with the `entity_resolution` and `graph_hop_retrieval` eval suites, and the published scale envelope.
 - `v0.8.0`: released. The memory-frontier waves — budgeted context packs, typed retrieval, staleness v1, agentic write protocol on the core MCP surface (11 tools), memory-quality eval suites, real scopes with key binding, merging consolidation, temporal slice, LongMemEval harness — plus Alice's first published benchmark result.
 - `v0.7.0`: released. Zero-infrastructure SQLite on-ramp — `uvx alice-memory mcp` serves the core tools with no Docker or Postgres; published to PyPI via Trusted Publishing.
@@ -14,7 +14,8 @@ In rough priority order:
 1. **Multi-session synthesis** — the weakest LongMemEval category (45.1%): retrieval breadth across sessions plus consolidation-driven aggregation; benchmark-driven iteration with the per-type breakdown as the scoreboard. The breadth ablation (49.2% at 2× context) motivates the planned query-shape-aware aggregation mode.
 2. **Dogfood daily** — run the stack against real agents with an embedding endpoint; calibrate the paraphrase target of the `retrieval_quality` benchmark; generate the usage telemetry that future ranking/policy improvements need.
 3. **Reference integrations** — deeper examples for popular agent frameworks on the core tool surface.
-4. **Hosted offering exploration** — the RLS posture and auth work make this plausible; still exploratory.
+4. **SQLite vector search at scale** — push the local-first on-ramp's embedding search past its documented ~20–30k comfort zone (brute-force cosine hits ~2.3s at 100k memories in the [scale envelope](docs/benchmarks/scale/README.md)): quantize stored embeddings (int8/binary) and rescore the top candidates at full precision, with the recall deltas published alongside the latency wins. Proven techniques, no new dependencies; recent results like Google's TurboQuant line (ICLR 2026) show how much headroom few-bit vector quantization has, and serve as the reference point for how far to take it.
+5. **Hosted offering exploration** — the RLS posture and auth work make this plausible; still exploratory.
 
 ## Explicit Non-Goals For Now
 - Hosted service / SLA commitments.