diff --git a/Gradata/src/gradata/_migrations/__init__.py b/Gradata/src/gradata/_migrations/__init__.py
index eddd13bb..148364da 100644
--- a/Gradata/src/gradata/_migrations/__init__.py
+++ b/Gradata/src/gradata/_migrations/__init__.py
@@ -72,6 +72,15 @@
         timestamp TEXT NOT NULL,
         user_context TEXT
     )""",
+    """CREATE TABLE IF NOT EXISTS sync_queue (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        payload_json TEXT NOT NULL,
+        kind TEXT NOT NULL CHECK(kind IN ('correction','lesson','event')),
+        enqueued_at REAL NOT NULL,
+        synced_at REAL,
+        attempts INTEGER DEFAULT 0,
+        last_error TEXT
+    )""",
 ]
 
 _MIGRATIONS: list[str] = [
@@ -97,6 +106,7 @@
     "ALTER TABLE super_meta_rules ADD COLUMN applies_when TEXT",
     "ALTER TABLE super_meta_rules ADD COLUMN never_when TEXT",
     "ALTER TABLE super_meta_rules ADD COLUMN transfer_scope TEXT DEFAULT 'personal'",
+    "CREATE INDEX IF NOT EXISTS idx_sync_queue_pending ON sync_queue(synced_at) WHERE synced_at IS NULL",
 ]
 
 
diff --git a/Gradata/src/gradata/_sync_queue.py b/Gradata/src/gradata/_sync_queue.py
new file mode 100644
index 00000000..2d82a0fe
--- /dev/null
+++ b/Gradata/src/gradata/_sync_queue.py
@@ -0,0 +1,125 @@
+"""Sync queue primitives — local SQLite buffer for write-through cloud sync.
+
+Day 1 of #194 (write-through sync). Provides CRUD over the ``sync_queue``
+table that is created as part of the standard Brain schema (see
+``_migrations/__init__.py``). The queue stores serialized payloads that
+need to be flushed to the cloud ingest endpoint; the actual flush
+worker is implemented later in the rollout.
+
+All public functions take an explicit ``sqlite3.Connection`` as the first
+argument — there are no implicit globals or module-level state.
+"""
+
+from __future__ import annotations
+
+import json
+import sqlite3
+import time
+
+__all__ = ["enqueue", "peek_pending", "mark_synced", "mark_failed"]
+
+_VALID_KINDS = ("correction", "lesson", "event")
+_MAX_ERROR_LEN = 500
+
+
+def enqueue(conn: sqlite3.Connection, kind: str, payload: dict) -> int:
+    """Insert a new row into ``sync_queue`` and return its row id.
+
+    Args:
+        conn: Active SQLite connection to a brain database.
+        kind: One of ``'correction'``, ``'lesson'``, ``'event'``. The
+            CHECK constraint on the column will reject anything else
+            with :class:`sqlite3.IntegrityError`.
+        payload: Arbitrary JSON-serializable dict; stored via
+            ``json.dumps`` into ``payload_json``.
+
+    Returns:
+        The autoincrement ``id`` of the newly inserted row.
+    """
+    payload_json = json.dumps(payload, ensure_ascii=False)
+    cur = conn.execute(
+        "INSERT INTO sync_queue (payload_json, kind, enqueued_at) "
+        "VALUES (?, ?, ?)",
+        (payload_json, kind, time.time()),
+    )
+    conn.commit()
+    rowid = cur.lastrowid
+    if rowid is None:
+        raise RuntimeError("sync_queue insert did not return a lastrowid")
+    return int(rowid)
+
+
+def peek_pending(conn: sqlite3.Connection, limit: int = 100) -> list[dict]:
+    """Return up to ``limit`` pending rows ordered by id (FIFO).
+
+    Pending == ``synced_at IS NULL``. Each returned dict has keys:
+    ``id`` (int), ``kind`` (str), ``payload`` (decoded dict),
+    ``enqueued_at`` (float), and ``attempts`` (int).
+    """
+    rows = conn.execute(
+        "SELECT id, kind, payload_json, enqueued_at, attempts "
+        "FROM sync_queue "
+        "WHERE synced_at IS NULL "
+        "ORDER BY id ASC "
+        "LIMIT ?",
+        (int(limit),),
+    ).fetchall()
+
+    out: list[dict] = []
+    for r in rows:
+        # Support both sqlite3.Row (named) and tuple rows.
+        try:
+            rid = r["id"]
+            kind = r["kind"]
+            payload_json = r["payload_json"]
+            enqueued_at = r["enqueued_at"]
+            attempts = r["attempts"]
+        except (TypeError, IndexError):
+            rid, kind, payload_json, enqueued_at, attempts = r
+        try:
+            payload = json.loads(payload_json) if payload_json else {}
+        except json.JSONDecodeError:
+            payload = {"_raw": payload_json}
+        out.append(
+            {
+                "id": int(rid),
+                "kind": kind,
+                "payload": payload,
+                "enqueued_at": float(enqueued_at) if enqueued_at is not None else None,
+                "attempts": int(attempts) if attempts is not None else 0,
+            }
+        )
+    return out
+
+
+def mark_synced(conn: sqlite3.Connection, ids: list[int]) -> None:
+    """Mark the given row ids as successfully synced (sets ``synced_at``)."""
+    if not ids:
+        return
+    now = time.time()
+    placeholders = ",".join("?" for _ in ids)
+    conn.execute(
+        f"UPDATE sync_queue SET synced_at = ? WHERE id IN ({placeholders})",
+        (now, *[int(i) for i in ids]),
+    )
+    conn.commit()
+
+
+def mark_failed(conn: sqlite3.Connection, ids: list[int], error: str) -> None:
+    """Record a failed sync attempt for the given rows.
+
+    Increments ``attempts`` and stores the (truncated) error message on
+    every targeted row. The rows remain pending (``synced_at`` stays
+    NULL) so they will be retried.
+    """
+    if not ids:
+        return
+    truncated = (error or "")[:_MAX_ERROR_LEN]
+    placeholders = ",".join("?" for _ in ids)
+    conn.execute(
+        f"UPDATE sync_queue "
+        f"SET attempts = COALESCE(attempts, 0) + 1, last_error = ? "
+        f"WHERE id IN ({placeholders})",
+        (truncated, *[int(i) for i in ids]),
+    )
+    conn.commit()
diff --git a/Gradata/src/gradata/onboard.py b/Gradata/src/gradata/onboard.py
index bd102085..0ec2abd8 100644
--- a/Gradata/src/gradata/onboard.py
+++ b/Gradata/src/gradata/onboard.py
@@ -103,6 +103,18 @@
         mention_count INTEGER DEFAULT 1
     )
     """,
+    # Write-through sync queue (cloud upload buffer, #194)
+    """
+    CREATE TABLE IF NOT EXISTS sync_queue (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        payload_json TEXT NOT NULL,
+        kind TEXT NOT NULL CHECK(kind IN ('correction','lesson','event')),
+        enqueued_at REAL NOT NULL,
+        synced_at REAL,
+        attempts INTEGER DEFAULT 0,
+        last_error TEXT
+    )
+    """,
 ]
 
 _INDEXES_SQL = [
@@ -113,6 +125,7 @@
     "CREATE INDEX IF NOT EXISTS idx_lesson_apps_lesson ON lesson_applications(lesson_id)",
     "CREATE INDEX IF NOT EXISTS idx_lesson_transitions_cat ON lesson_transitions(category)",
     "CREATE INDEX IF NOT EXISTS idx_lesson_transitions_desc ON lesson_transitions(lesson_desc)",
+    "CREATE INDEX IF NOT EXISTS idx_sync_queue_pending ON sync_queue(synced_at) WHERE synced_at IS NULL",
 ]
 
 # ── Subdirectories ────────────────────────────────────────────────────
diff --git a/Gradata/tests/test_sync_queue.py b/Gradata/tests/test_sync_queue.py
new file mode 100644
index 00000000..eb00bec8
--- /dev/null
+++ b/Gradata/tests/test_sync_queue.py
@@ -0,0 +1,129 @@
+"""Unit tests for sync_queue primitives (#194, day 1).
+
+Covers schema creation via the brain migration module plus CRUD over the
+``sync_queue`` table through :mod:`gradata._sync_queue`.
+"""
+
+from __future__ import annotations
+
+import sqlite3
+import time
+
+import pytest
+
+from gradata import _sync_queue
+from gradata._migrations import _BASE_TABLES, _MIGRATIONS
+
+
+def _make_conn() -> sqlite3.Connection:
+    """Return an in-memory SQLite connection with the sync_queue schema applied.
+
+    Pulls the canonical DDL straight from ``gradata._migrations`` so we
+    are testing the same schema a real brain would get.
+    """
+    conn = sqlite3.connect(":memory:")
+    conn.row_factory = sqlite3.Row
+    for sql in _BASE_TABLES:
+        if "sync_queue" in sql:
+            conn.execute(sql)
+    for sql in _MIGRATIONS:
+        if "sync_queue" in sql:
+            conn.execute(sql)
+    conn.commit()
+    return conn
+
+
+def test_sync_queue_enqueue_returns_int_and_peek_finds_row():
+    conn = _make_conn()
+    payload = {"draft": "hello", "final": "Hello.", "meta": {"n": 1}}
+
+    row_id = _sync_queue.enqueue(conn, "correction", payload)
+
+    assert isinstance(row_id, int)
+    assert row_id > 0
+
+    pending = _sync_queue.peek_pending(conn)
+    assert len(pending) == 1
+    row = pending[0]
+    assert row["id"] == row_id
+    assert row["kind"] == "correction"
+    # Payload roundtrips exactly through json.dumps/loads.
+    assert row["payload"] == payload
+    assert row["attempts"] == 0
+    assert isinstance(row["enqueued_at"], float)
+
+
+def test_sync_queue_peek_pending_respects_limit_and_fifo_order():
+    conn = _make_conn()
+    ids = [_sync_queue.enqueue(conn, "event", {"i": i}) for i in range(5)]
+
+    pending = _sync_queue.peek_pending(conn, limit=3)
+    assert [r["id"] for r in pending] == ids[:3]
+
+
+def test_sync_queue_mark_synced_removes_from_peek_pending():
+    conn = _make_conn()
+    id_a = _sync_queue.enqueue(conn, "correction", {"x": 1})
+    id_b = _sync_queue.enqueue(conn, "lesson", {"x": 2})
+
+    _sync_queue.mark_synced(conn, [id_a])
+
+    pending_ids = [r["id"] for r in _sync_queue.peek_pending(conn)]
+    assert id_a not in pending_ids
+    assert id_b in pending_ids
+
+    # synced_at should be a float timestamp on row A.
+    row = conn.execute(
+        "SELECT synced_at FROM sync_queue WHERE id = ?", (id_a,)
+    ).fetchone()
+    assert row["synced_at"] is not None
+    assert float(row["synced_at"]) <= time.time() + 1.0
+
+
+def test_sync_queue_mark_failed_increments_attempts_and_keeps_row_pending():
+    conn = _make_conn()
+    row_id = _sync_queue.enqueue(conn, "event", {"k": "v"})
+
+    _sync_queue.mark_failed(conn, [row_id], "boom: network unreachable")
+    _sync_queue.mark_failed(conn, [row_id], "boom again")
+
+    pending = _sync_queue.peek_pending(conn)
+    assert len(pending) == 1
+    assert pending[0]["id"] == row_id
+    assert pending[0]["attempts"] == 2
+
+    row = conn.execute(
+        "SELECT last_error, synced_at FROM sync_queue WHERE id = ?", (row_id,)
+    ).fetchone()
+    assert row["last_error"] == "boom again"
+    assert row["synced_at"] is None
+
+
+def test_sync_queue_mark_failed_truncates_long_error_to_500_chars():
+    conn = _make_conn()
+    row_id = _sync_queue.enqueue(conn, "event", {})
+    long_err = "x" * 5000
+
+    _sync_queue.mark_failed(conn, [row_id], long_err)
+
+    row = conn.execute(
+        "SELECT last_error FROM sync_queue WHERE id = ?", (row_id,)
+    ).fetchone()
+    assert len(row["last_error"]) == 500
+    assert row["last_error"] == "x" * 500
+
+
+def test_sync_queue_kind_constraint_rejects_invalid_value():
+    conn = _make_conn()
+    with pytest.raises(sqlite3.IntegrityError):
+        _sync_queue.enqueue(conn, "foo", {"bad": True})
+
+
+def test_sync_queue_empty_id_lists_are_noops():
+    conn = _make_conn()
+    # Should not raise, should not touch the DB.
+    _sync_queue.mark_synced(conn, [])
+    _sync_queue.mark_failed(conn, [], "nope")
+
+    pending = _sync_queue.peek_pending(conn)
+    assert pending == []