diff --git a/CHANGELOG.md b/CHANGELOG.md
index c0a42c2..b91b45a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -23,6 +23,7 @@ the first consumer-visible behaviour change and will drive the next SDK version
 - **Automation-detector wording**: the anti-tamper flag now reads "Anti-tamper signals" (not "Automation detected") when the combined confidence is weak — e.g. devtools open in a dev environment — so a human isn't labelled a bot. The machine-readable `code` (`automation_suspected`) is unchanged; the reason text also drops the `tamper.` prefix for readability.
 
 ### Internal
+- **Retention sweeper** ([ADR-0004](docs/adr/0004-consent-and-data-lifecycle.md)): a daily BullMQ repeatable job in the worker deletes identities (and, by cascade, their snapshots/drifts/risk/links) whose `last_seen` is older than their project's `retention_days`; projects with a null `retention_days` keep data indefinitely. `sweepRetention()` is idempotent and unit-tested against a real DB.
 - **Data-subject endpoints** ([ADR-0004](docs/adr/0004-consent-and-data-lifecycle.md)): `DELETE /v1/identity/:id` (GDPR Art. 17 — erases the identity; snapshots/drifts/risk assessments/account links cascade) and `GET /v1/identity/:id/export` (Art. 20 — the full bundle, including each snapshot's consent provenance). Erasure is strictly key-gated (a non-GET never reaches the route via an admin session); export is readable by key or admin session like the other reads.
 - **Consent provenance + IP minimization on the server** ([ADR-0004](docs/adr/0004-consent-and-data-lifecycle.md), migration 012): snapshots now record the lawful basis, consent version, and grant time the SDK forwards (falling back to a per-project `lawful_basis_default`). The stored `client_ip` is **network-truncated by default** (`/24` IPv4, `/48` IPv6 via `minimizeIp`) — still city-accurate for impossible-travel, while the full IP is used only transiently in the request path (so the anonymizer detector is unaffected). A per-project `store_full_ip` flag keeps the full address for operators with a documented basis. New `projects.retention_days` column lands here for the upcoming retention sweeper.
 - **Demo recognises a new device on re-observe**: the demo showed the synchronous `POST /v1/resolve` result, then committed via `sdk.flush()` (async ingest). `flush()` returns on enqueue, not on commit, so a first-sight device read as "new" on every click. The demo now polls resolve (bounded) until the commit lands before re-enabling Observe, and the status spells out the eventual-consistency step — so the next click resolves against committed history and recognises the device.
diff --git a/packages/server/src/pipeline/retention.integration.test.ts b/packages/server/src/pipeline/retention.integration.test.ts
new file mode 100644
index 0000000..6402952
--- /dev/null
+++ b/packages/server/src/pipeline/retention.integration.test.ts
@@ -0,0 +1,68 @@
+import { describe, it, expect, beforeAll, afterAll } from 'vitest';
+import { migrate } from '../db/migrate.js';
+import { db } from '../db/client.js';
+import { sweepRetention } from './retention.js';
+import { hashApiKey } from '../middleware/api-key.js';
+
+// Gated on DATABASE_URL like the other integration suites (runs in CI, skips locally
+// without a DB). See events.integration.test.ts for the rationale.
+const hasDb = Boolean(process.env['DATABASE_URL']);
+const API_KEY = 'retention-test-key';
+
+let projectId: string; // retention_days = 30
+
+async function seedIdentity(project: string, ageDays: number): Promise<string> {
+  const id = `ret-${ageDays}d-${crypto.randomUUID()}`;
+  await db`
+    INSERT INTO identities (id, project_id, last_seen)
+    VALUES (${id}, ${project}, now() - make_interval(days => ${ageDays}))
+  `;
+  return id;
+}
+
+beforeAll(async () => {
+  if (!hasDb) return;
+  await migrate();
+  await db`DELETE FROM projects WHERE api_key_hash IN (${hashApiKey(API_KEY)}, ${hashApiKey(`${API_KEY}-keep`)})`;
+  const [proj] = await db<{ id: string }[]>`
+    INSERT INTO projects (api_key_hash, name, retention_days)
+    VALUES (${hashApiKey(API_KEY)}, 'Retention Test', 30) RETURNING id
+  `;
+  projectId = proj!.id;
+});
+
+afterAll(async () => {
+  if (!hasDb) return;
+  await db`DELETE FROM projects WHERE api_key_hash IN (${hashApiKey(API_KEY)}, ${hashApiKey(`${API_KEY}-keep`)})`;
+  await db.end();
+});
+
+describe.skipIf(!hasDb)('sweepRetention', () => {
+  it('deletes identities older than retention_days and keeps recent ones', async () => {
+    const stale = await seedIdentity(projectId, 60); // older than 30d → swept
+    const fresh = await seedIdentity(projectId, 1); // within 30d → kept
+
+    const result = await sweepRetention(db);
+    expect(result.identitiesDeleted).toBeGreaterThanOrEqual(1);
+
+    const rows = await db<{ id: string }[]>`
+      SELECT id FROM identities WHERE project_id = ${projectId}
+    `;
+    const ids = rows.map((r) => r.id);
+    expect(ids).toContain(fresh);
+    expect(ids).not.toContain(stale);
+  });
+
+  it('skips projects with null retention_days (keep forever)', async () => {
+    const [keepProj] = await db<{ id: string }[]>`
+      INSERT INTO projects (api_key_hash, name)
+      VALUES (${hashApiKey(`${API_KEY}-keep`)}, 'Keep Forever') RETURNING id
+    `;
+    const ancient = await seedIdentity(keepProj!.id, 999);
+
+    await sweepRetention(db);
+
+    const rows = await db<{ id: string }[]>`SELECT id FROM identities WHERE id = ${ancient}`;
+    expect(rows.length).toBe(1); // untouched — retention_days is null
+  });
+});
diff --git a/packages/server/src/pipeline/retention.ts b/packages/server/src/pipeline/retention.ts
new file mode 100644
index 0000000..dee19e2
--- /dev/null
+++ b/packages/server/src/pipeline/retention.ts
@@ -0,0 +1,35 @@
+import type { Sql } from 'postgres';
+import { logger } from '../logger.js';
+
+// The repeatable retention sweep runs on its own BullMQ queue (separate from ingest).
+export const RETENTION_QUEUE_NAME = 'retention';
+
+// Deletes identities whose most recent activity (last_seen) is older than their
+// project's retention_days; snapshots, drifts, risk assessments, cluster merges, and
+// account links cascade (ON DELETE CASCADE on identity_id). Projects with a null or
+// non-positive retention_days keep data indefinitely and are skipped. Idempotent —
+// safe to run repeatedly. Returns per-run counts for logging. (GDPR data-lifecycle,
+// ADR-0004.)
+export async function sweepRetention(
+  sql: Sql,
+): Promise<{ projects: number; identitiesDeleted: number }> {
+  const projects = await sql<{ id: string; retention_days: number }[]>`
+    SELECT id, retention_days
+    FROM projects
+    WHERE retention_days IS NOT NULL AND retention_days > 0
+  `;
+
+  let identitiesDeleted = 0;
+  for (const p of projects) {
+    const deleted = await sql<{ id: string }[]>`
+      DELETE FROM identities
+      WHERE project_id = ${p.id}
+        AND last_seen < now() - make_interval(days => ${p.retention_days})
+      RETURNING id
+    `;
+    identitiesDeleted += deleted.length;
+  }
+
+  logger.info({ projects: projects.length, identitiesDeleted }, 'retention sweep complete');
+  return { projects: projects.length, identitiesDeleted };
+}
diff --git a/packages/server/src/worker.ts b/packages/server/src/worker.ts
index 180896b..0319282 100644
--- a/packages/server/src/worker.ts
+++ b/packages/server/src/worker.ts
@@ -1,10 +1,11 @@
 import { startTracing } from './tracing.js';
 startTracing();
 
-import { Worker } from 'bullmq';
+import { Queue, Worker } from 'bullmq';
 import { createQueueConnection, INGEST_QUEUE_NAME } from './queue/ingest.js';
 import type { IngestJobData } from './queue/ingest.js';
 import { resolveSnapshot } from './pipeline/resolve.js';
+import { sweepRetention, RETENTION_QUEUE_NAME } from './pipeline/retention.js';
 import { db } from './db/client.js';
 // Imported after ./tracing.js so the OTel pino instrumentation patches pino and
 // injects trace_id/span_id into log lines — same ordering rationale as index.ts.
@@ -36,11 +37,30 @@ worker.on('failed', (job, err) => {
   );
 });
 
+// Daily retention sweep (GDPR data-lifecycle, ADR-0004). A repeatable job is enqueued
+// idempotently (BullMQ dedupes by repeat key, so re-adding on every boot is safe) and
+// processed by a dedicated worker that deletes data past each project's retention_days.
+const retentionQueue = new Queue(RETENTION_QUEUE_NAME, { connection: createQueueConnection() });
+void retentionQueue
+  .add('sweep', {}, { repeat: { pattern: '30 3 * * *' }, removeOnComplete: true, removeOnFail: 50 })
+  .catch((err) => logger.error({ err }, 'failed to schedule retention sweep'));
+
+const retentionWorker = new Worker(
+  RETENTION_QUEUE_NAME,
+  async () => sweepRetention(db),
+  { connection: createQueueConnection() },
+);
+retentionWorker.on('failed', (job, err) => {
+  logger.error({ jobId: job?.id, err }, 'retention sweep failed');
+});
+
 // Drain in-flight jobs and close the DB pool on shutdown so a redeploy doesn't drop
 // work mid-resolution.
 async function shutdown(): Promise<void> {
   logger.info('scent-worker shutting down');
   await worker.close();
+  await retentionWorker.close();
+  await retentionQueue.close();
   await db.end();
   process.exit(0);
 }