diff --git a/CHANGELOG.md b/CHANGELOG.md
index 74a6d475d4..38a69715de 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,6 +6,7 @@
 
 ### Added
 
+- `bench/session-memory.bench.ts` reports heap after a forced GC, current RSS and high-water RSS at each of three phases (module baseline, `SessionManager.open`, `buildSessionContext`) over a synthetic transcript sized by `SESSION_MB`.
 - Added `model.toolCallLoopGuard.readSubsumptionThreshold` (default `2`) to steer models that re-read unchanged code lines back-to-back before consuming full context.
 - `VEYYON_DEBUG_STARTUP=1` writes one line per phase of a prompt submission (compaction check, plan arm, context build, memory context), so a slow submit names the phase that spent the time.
 - `read` takes `depth` and `limit` arguments for directory listings, and a read of the session working directory root with neither now returns a concise top-level listing with per-subdirectory entry counts instead of the recursive tree.
diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
index e6eafd13b0..bcfc760398 100644
--- a/packages/coding-agent/CHANGELOG.md
+++ b/packages/coding-agent/CHANGELOG.md
@@ -7,6 +7,8 @@
 - Classified runner output (cargo, bun, Go, ctest, dotnet, clippy, golangci-lint, Gradle lint, pytest, and tsc/eslint-family) now opens with a result-contract header: `[clean] <command>` or `[errors]` / `[errors N] <command>`. The header is the verdict and the body contains retained diagnostics.
 ### Added
 
+- `bench/session-memory.bench.ts` reports heap after a forced GC, current RSS and high-water RSS at each of three phases (module baseline, `SessionManager.open`, `buildSessionContext`) over a synthetic transcript sized by `SESSION_MB`.
+
 - Added `model.toolCallLoopGuard.readSubsumptionThreshold` (default `2`) to steer models that re-read unchanged code lines back-to-back before consuming full context.
 - `VEYYON_DEBUG_STARTUP=1` writes one line per phase of a prompt submission (compaction check, plan arm, context build, memory context), so a slow submit names the phase that spent the time.
 - `read` takes `depth` and `limit` arguments for directory listings, and a read of the session working directory root with neither now returns a concise top-level listing with per-subdirectory entry counts instead of the recursive tree.
diff --git a/packages/coding-agent/bench/session-memory.bench.ts b/packages/coding-agent/bench/session-memory.bench.ts
new file mode 100644
index 0000000000..4e780a97c3
--- /dev/null
+++ b/packages/coding-agent/bench/session-memory.bench.ts
@@ -0,0 +1,128 @@
+/**
+ * Benchmark: resident memory of loading a large session through the real
+ * pipeline.
+ *
+ * Reports three numbers per phase: the heap AFTER a forced synchronous GC,
+ * which is steady-state retention; the current RSS; and the process high-water
+ * RSS, which is monotonic, so the rise between two phases is the transient peak
+ * that phase reached.
+ *
+ *   1. module baseline      — before any session work
+ *   2. entries loaded       — retention of the parsed entry graph
+ *   3. context build        — buildSessionContext over the loaded branch
+ *
+ * The synthetic transcript mirrors real sessions: alternating user /
+ * assistant turns with multi-KB text blocks and periodic large tool outputs.
+ *
+ * Run: bun packages/coding-agent/bench/session-memory.bench.ts
+ *   SESSION_MB=80 bun ... to size the synthetic transcript.
+ */
+
+import * as fs from "node:fs";
+import * as path from "node:path";
+import { SessionManager } from "../src/session/session-manager";
+
+const TARGET_MB = Number(process.env.SESSION_MB ?? "32");
+const TOOL_OUTPUT_EVERY = 10;
+const TEXT_KB = 4;
+
+function makeEntry(i: number, big: string): string {
+	const timestamp = new Date(Date.UTC(2026, 0, 1) + i * 1000).toISOString();
+	const base = { type: "message", id: `e${i}`, parentId: i === 0 ? null : `e${i - 1}`, timestamp };
+	if (i % TOOL_OUTPUT_EVERY === TOOL_OUTPUT_EVERY - 1) {
+		return JSON.stringify({
+			...base,
+			message: { role: "toolResult", toolCallId: `tc${i}`, toolName: "bash", output: big },
+		});
+	}
+	if (i % 2 === 0) {
+		return JSON.stringify({
+			...base,
+			message: { role: "user", content: [{ type: "text", text: `Task step ${i}: ${big.slice(0, 400)}` }] },
+		});
+	}
+	return JSON.stringify({
+		...base,
+		message: {
+			role: "assistant",
+			content: [{ type: "text", text: `Analysis ${i}: ${big}` }],
+			api: "openai-completions",
+			provider: "openai",
+			model: "gpt-test",
+			usage: { input: 1000, output: 2000, cacheRead: 0, cacheWrite: 0, totalTokens: 3000, cost: { total: 0.01 } },
+			stopReason: "stop",
+		},
+	});
+}
+
+function writeSyntheticSession(dir: string): { file: string; bytes: number; count: number } {
+	const file = path.join(dir, "memory-bench.jsonl");
+	const big = "x".repeat(TEXT_KB * 1024);
+	let written = 0;
+	let count = 0;
+	const fd = fs.openSync(file, "w");
+	fs.writeSync(
+		fd,
+		`${JSON.stringify({ type: "session", version: 3, id: "membench0000000000000000000000", timestamp: new Date().toISOString(), cwd: dir })}\n`,
+	);
+	while (written < TARGET_MB * 1024 * 1024) {
+		const line = `${makeEntry(count++, big)}\n`;
+		fs.writeSync(fd, line);
+		written += Buffer.byteLength(line);
+	}
+	fs.closeSync(fd);
+	return { file, bytes: written, count };
+}
+
+/** High-water RSS in MiB. Linux reads the kernel's own VmHWM accounting;
+ * other platforms fall back to the current RSS sampling. */
+function peakRssMiB(): number {
+	try {
+		const status = fs.readFileSync("/proc/self/status", "utf8");
+		const m = /VmHWM:\s+(\d+) kB/.exec(status);
+		if (m) return Number(m[1]) / 1024;
+	} catch {
+		// not Linux
+	}
+	return process.memoryUsage().rss / 1048576;
+}
+
+function gcAndReport(label: string): void {
+	// Bun.gc(true) has no portable equivalent: node's global.gc exists only under
+	// --expose-gc, which this script cannot set for its own process. Without a
+	// forced collection the heap number is whatever the collector last felt like
+	// doing, which is not retention.
+	Bun.gc(true);
+	const heap = process.memoryUsage().heapUsed / 1048576;
+	const rss = process.memoryUsage().rss / 1048576;
+	const peak = peakRssMiB();
+	console.log(
+		`${label.padEnd(28)} heap ${heap.toFixed(0).padStart(5)} MiB   rss ${rss.toFixed(0).padStart(5)} MiB   peak ${peak.toFixed(0).padStart(5)} MiB`,
+	);
+}
+
+// The transcript is written under the repo's gitignored scratch root rather than
+// the system temp dir: at SESSION_MB=500 this is half a gigabyte, and on a host
+// whose /tmp is a tmpfs that is half a gigabyte of RAM charged against the
+// measurement the bench exists to take.
+const scratchRoot = path.join(import.meta.dirname, "..", "..", "..", ".scratch");
+fs.mkdirSync(scratchRoot, { recursive: true });
+const dir = fs.mkdtempSync(path.join(scratchRoot, "session-memory-"));
+const { file, bytes, count } = writeSyntheticSession(dir);
+console.log(`synthetic session: ${(bytes / 1048576).toFixed(1)} MiB, ${count} entries (${TARGET_MB} MiB target)`);
+gcAndReport("module baseline");
+
+const t0 = performance.now();
+const sm = await SessionManager.open(file, undefined, undefined, { suppressBreadcrumb: true });
+const loadMs = performance.now() - t0;
+gcAndReport(`entries loaded (${loadMs.toFixed(0)}ms)`);
+
+const t1 = performance.now();
+// The manager's own accessor, not the free function: it passes the live entry
+// array, leaf and id index, which is the shape the production build sees.
+const context = sm.buildSessionContext();
+const buildMs = performance.now() - t1;
+gcAndReport(`context build (${buildMs.toFixed(0)}ms)`);
+console.log(`context messages             ${context.messages.length}`);
+
+fs.rmSync(dir, { recursive: true, force: true });