diff --git a/AGENTS.md b/AGENTS.md
index 935f35f91e4..a91c21cf601 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -93,6 +93,7 @@
 - Naming: match source names with `*.test.ts`; e2e in `*.e2e.test.ts`.
 - Run `pnpm test` (or `pnpm test:coverage`) before pushing when you touch logic.
 - Do not set test workers above 16; tried already.
+- If local Vitest runs cause memory pressure (common on non-Mac-Studio hosts), use `OPENCLAW_TEST_PROFILE=low OPENCLAW_TEST_SERIAL_GATEWAY=1 pnpm test` for land/gate runs.
 - Live tests (real keys): `CLAWDBOT_LIVE_TEST=1 pnpm test:live` (OpenClaw-only) or `LIVE=1 pnpm test:live` (includes provider live tests). Docker: `pnpm test:docker:live-models`, `pnpm test:docker:live-gateway`. Onboarding Docker E2E: `pnpm test:docker:onboard`.
 - Full kit + what’s covered: `docs/testing.md`.
 - Changelog: user-facing changes only; no internal/meta notes (version alignment, appcast reminders, release process).
diff --git a/scripts/test-parallel.mjs b/scripts/test-parallel.mjs
index 94d8c0726dd..cb7e950a5da 100644
--- a/scripts/test-parallel.mjs
+++ b/scripts/test-parallel.mjs
@@ -157,16 +157,21 @@ const testProfile =
 const overrideWorkers = Number.parseInt(process.env.OPENCLAW_TEST_WORKERS ?? "", 10);
 const resolvedOverride =
   Number.isFinite(overrideWorkers) && overrideWorkers > 0 ? overrideWorkers : null;
-// Keep gateway serial on Windows CI and CI by default; run in parallel locally
-// for lower wall-clock time. CI can opt in via OPENCLAW_TEST_PARALLEL_GATEWAY=1.
+const hostCpuCount = os.cpus().length;
+const hostMemoryGiB = Math.floor(os.totalmem() / 1024 ** 3);
+// Keep aggressive local defaults for high-memory workstations (Mac Studio class).
+const highMemLocalHost = !isCI && hostMemoryGiB >= 96;
+const lowMemLocalHost = !isCI && hostMemoryGiB < 64;
+const parallelGatewayEnabled =
+  process.env.OPENCLAW_TEST_PARALLEL_GATEWAY === "1" || (!isCI && highMemLocalHost);
+// Keep gateway serial by default except when explicitly requested or on high-memory local hosts.
 const keepGatewaySerial =
   isWindowsCi ||
   process.env.OPENCLAW_TEST_SERIAL_GATEWAY === "1" ||
   testProfile === "serial" ||
-  (isCI && process.env.OPENCLAW_TEST_PARALLEL_GATEWAY !== "1");
+  !parallelGatewayEnabled;
 const parallelRuns = keepGatewaySerial ? runs.filter((entry) => entry.name !== "gateway") : runs;
 const serialRuns = keepGatewaySerial ? runs.filter((entry) => entry.name === "gateway") : [];
-const hostCpuCount = os.cpus().length;
 const baseLocalWorkers = Math.max(4, Math.min(16, hostCpuCount));
 const loadAwareDisabledRaw = process.env.OPENCLAW_TEST_LOAD_AWARE?.trim().toLowerCase();
 const loadAwareDisabled = loadAwareDisabledRaw === "0" || loadAwareDisabledRaw === "false";
@@ -199,15 +204,29 @@ const defaultWorkerBudget =
             extensions: Math.max(1, Math.min(6, Math.floor(localWorkers / 2))),
             gateway: Math.max(1, Math.min(2, Math.floor(localWorkers / 4))),
           }
-        : {
-            // Local `pnpm test` runs multiple vitest groups concurrently;
-            // bias workers toward unit-fast (wall-clock bottleneck) while
-            // keeping unit-isolated low enough that both groups finish closer together.
-            unit: Math.max(4, Math.min(14, Math.floor((localWorkers * 7) / 8))),
-            unitIsolated: Math.max(1, Math.min(2, Math.floor(localWorkers / 6) || 1)),
-            extensions: Math.max(1, Math.min(4, Math.floor(localWorkers / 4))),
-            gateway: Math.max(2, Math.min(4, Math.floor(localWorkers / 3))),
-          };
+        : highMemLocalHost
+          ? {
+              // High-memory local hosts can prioritize wall-clock speed.
+              unit: Math.max(4, Math.min(14, Math.floor((localWorkers * 7) / 8))),
+              unitIsolated: Math.max(1, Math.min(2, Math.floor(localWorkers / 6) || 1)),
+              extensions: Math.max(1, Math.min(4, Math.floor(localWorkers / 4))),
+              gateway: Math.max(2, Math.min(4, Math.floor(localWorkers / 3))),
+            }
+          : lowMemLocalHost
+            ? {
+                // Sub-64 GiB local hosts are prone to OOM with large vmFork runs.
+                unit: 2,
+                unitIsolated: 1,
+                extensions: 1,
+                gateway: 1,
+              }
+            : {
+                // 64-95 GiB local hosts: conservative split with some parallel headroom.
+                unit: Math.max(2, Math.min(8, Math.floor(localWorkers / 2))),
+                unitIsolated: 1,
+                extensions: Math.max(1, Math.min(4, Math.floor(localWorkers / 4))),
+                gateway: 1,
+              };
 
 // Keep worker counts predictable for local runs; trim macOS CI workers to avoid worker crashes/OOM.
 // In CI on linux/windows, prefer Vitest defaults to avoid cross-test interference from lower worker counts.