adk-memshield 0.1.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,16 @@
1
+ """adk-memshield: Zero-trust multi-tenant semantic memory service for Google ADK."""
2
+
3
+ from .config import MemShieldConfig
4
+ from .embedder import EmbedFn, MemShieldEmbedder
5
+ from .service import MemShieldMemoryService, QueryTelemetry, register_adk_uri
6
+
7
+ __version__ = "0.1.0"
8
+
9
+ __all__ = [
10
+ "EmbedFn",
11
+ "MemShieldConfig",
12
+ "MemShieldEmbedder",
13
+ "MemShieldMemoryService",
14
+ "QueryTelemetry",
15
+ "register_adk_uri",
16
+ ]
@@ -0,0 +1,62 @@
1
+ """Configuration models for adk-memshield."""
2
+
3
+ from __future__ import annotations
4
+
5
+ import os
6
+ import re
7
+ from typing import Literal
8
+
9
+ from pydantic import BaseModel, ConfigDict, Field, field_validator
10
+
11
+ _SAFE_IDENTIFIER_RE = re.compile(r"^[a-zA-Z_][a-zA-Z0-9_]*$")
12
+
13
+
14
+ class MemShieldConfig(BaseModel):
15
+ """Configuration for MemShieldMemoryService.
16
+
17
+ Attributes:
18
+ dsn: PostgreSQL connection URI using a non-superuser role (e.g. memshield_app)
19
+ so PostgreSQL Row-Level Security (RLS) policies are strictly enforced.
20
+ tenant_id: Active tenant identifier bound to transaction-local context
21
+ (`app.current_tenant_id`).
22
+ agent_clearance: Clearance level (1..5) of the active agent/caller bound to
23
+ transaction-local context (`app.agent_clearance`).
24
+ default_scope: Default visibility scope for newly stored memories
25
+ (`'private'` or `'tenant_shared'`).
26
+ default_min_clearance: Default minimum clearance level (1..5) required to
27
+ read `'tenant_shared'` memories.
28
+ top_k: Maximum number of memories returned by `search_memory`.
29
+ min_similarity: Optional minimum cosine similarity threshold for retrieval.
30
+ embedding_model: FastEmbed ONNX model identifier.
31
+ embedding_dim: Expected embedding vector dimensionality.
32
+ table_name: Target PostgreSQL table name (`'memories'` by default).
33
+ simulate_missing_app_filter: Chaos/testing flag. When True, omits application-layer
34
+ SQL `WHERE` clauses during `search_memory` to verify that PostgreSQL RLS
35
+ independently prevents cross-tenant and cross-user leakage.
36
+ """
37
+
38
+ model_config = ConfigDict(validate_assignment=True, extra="forbid")
39
+
40
+ dsn: str = Field(
41
+ default_factory=lambda: os.environ.get(
42
+ "MEMSHIELD_DB_DSN",
43
+ "postgresql://memshield_app:memshield_pass@localhost:54329/memshield",
44
+ )
45
+ )
46
+ tenant_id: str = Field(default="default", min_length=1)
47
+ agent_clearance: int = Field(default=1, ge=1, le=5)
48
+ default_scope: Literal["private", "tenant_shared"] = Field(default="private")
49
+ default_min_clearance: int = Field(default=1, ge=1, le=5)
50
+ top_k: int = Field(default=5, ge=1, le=100)
51
+ min_similarity: float | None = Field(default=None, ge=-1.0, le=1.0)
52
+ embedding_model: str = Field(default="BAAI/bge-small-en-v1.5")
53
+ embedding_dim: int = Field(default=384, ge=1)
54
+ table_name: str = Field(default="memories")
55
+ simulate_missing_app_filter: bool = Field(default=False)
56
+
57
+ @field_validator("table_name")
58
+ @classmethod
59
+ def _validate_table_name(cls, value: str) -> str:
60
+ if not _SAFE_IDENTIFIER_RE.match(value):
61
+ raise ValueError(f"Unsafe SQL table identifier: {value!r}")
62
+ return value
@@ -0,0 +1,65 @@
1
+ """FastEmbed ONNX embedding wrapper with custom callable override support."""
2
+
3
+ from __future__ import annotations
4
+
5
+ import threading
6
+ from collections.abc import Callable, Sequence
7
+ from typing import TYPE_CHECKING
8
+
9
+ if TYPE_CHECKING:
10
+ from fastembed import TextEmbedding
11
+
12
+ EmbedFn = Callable[[Sequence[str]], list[list[float]]]
13
+
14
+ _MODEL_CACHE: dict[str, TextEmbedding] = {}
15
+ _MODEL_CACHE_LOCK = threading.Lock()
16
+
17
+
18
+ def _get_fastembed_model(model_name: str) -> TextEmbedding:
19
+ """Return a process-cached FastEmbed TextEmbedding instance."""
20
+ with _MODEL_CACHE_LOCK:
21
+ if model_name not in _MODEL_CACHE:
22
+ from fastembed import TextEmbedding
23
+
24
+ _MODEL_CACHE[model_name] = TextEmbedding(model_name=model_name)
25
+ return _MODEL_CACHE[model_name]
26
+
27
+
28
+ class MemShieldEmbedder:
29
+ """Embeds text strings into normalized float vectors for pgvector storage and search.
30
+
31
+ Defaults to `fastembed` (`BAAI/bge-small-en-v1.5`, 384 dimensions) via ONNX Runtime,
32
+ or delegates to a user-supplied `embed_fn`.
33
+ """
34
+
35
+ def __init__(
36
+ self,
37
+ model_name: str = "BAAI/bge-small-en-v1.5",
38
+ expected_dim: int = 384,
39
+ embed_fn: EmbedFn | None = None,
40
+ ) -> None:
41
+ self.model_name = model_name
42
+ self.expected_dim = expected_dim
43
+ self._embed_fn = embed_fn
44
+
45
+ def embed_documents(self, texts: Sequence[str]) -> list[list[float]]:
46
+ """Embed a batch of document/memory strings into float lists."""
47
+ if not texts:
48
+ return []
49
+ if self._embed_fn is not None:
50
+ vectors = [list(map(float, vec)) for vec in self._embed_fn(texts)]
51
+ else:
52
+ model = _get_fastembed_model(self.model_name)
53
+ vectors = [[float(x) for x in vec.tolist()] for vec in model.embed(list(texts))]
54
+
55
+ for vec in vectors:
56
+ if len(vec) != self.expected_dim:
57
+ raise ValueError(
58
+ f"Embedding dimension mismatch: expected {self.expected_dim}, got {len(vec)}"
59
+ )
60
+ return vectors
61
+
62
+ def embed_query(self, query: str) -> list[float]:
63
+ """Embed a single search query string into a float list."""
64
+ vectors = self.embed_documents([query])
65
+ return vectors[0]
@@ -0,0 +1,598 @@
1
+ """Google ADK BaseMemoryService implementation backed by PostgreSQL pgvector + RLS."""
2
+
3
+ from __future__ import annotations
4
+
5
+ import asyncio
6
+ import json
7
+ import time
8
+ from collections.abc import Iterator, Mapping, Sequence
9
+ from contextlib import contextmanager
10
+ from dataclasses import asdict, dataclass
11
+ from typing import Any, Literal
12
+ from urllib.parse import parse_qs, urlencode, urlparse, urlunparse
13
+
14
+ import psycopg
15
+ from google.adk.events.event import Event
16
+ from google.adk.memory.base_memory_service import BaseMemoryService, SearchMemoryResponse
17
+ from google.adk.memory.memory_entry import MemoryEntry
18
+ from google.adk.sessions.session import Session
19
+ from google.genai import types
20
+ from psycopg_pool import ConnectionPool
21
+
22
+ from .config import MemShieldConfig
23
+ from .embedder import EmbedFn, MemShieldEmbedder
24
+
25
+ _SHARED_POOLS: dict[str, ConnectionPool] = {}
26
+
27
+
28
+ def _get_shared_pool(dsn: str) -> ConnectionPool:
29
+ """Return a process-shared ConnectionPool keyed by PostgreSQL DSN."""
30
+ pool = _SHARED_POOLS.get(dsn)
31
+ if pool is None or pool.closed:
32
+ pool = ConnectionPool(conninfo=dsn, min_size=1, max_size=8, open=True)
33
+ _SHARED_POOLS[dsn] = pool
34
+ return pool
35
+
36
+
37
+ def _format_vector(vec: Sequence[float]) -> str:
38
+ """Serialize a float sequence as a pgvector literal string."""
39
+ return "[" + ",".join(f"{float(x):.8f}" for x in vec) + "]"
40
+
41
+
42
+ def _extract_text_from_content(content: types.Content | None) -> str:
43
+ """Extract concatenated non-empty text parts from a google.genai Content object."""
44
+ if content is None or not content.parts:
45
+ return ""
46
+ texts: list[str] = []
47
+ for part in content.parts:
48
+ if getattr(part, "text", None):
49
+ stripped = part.text.strip()
50
+ if stripped:
51
+ texts.append(stripped)
52
+ return "\n".join(texts)
53
+
54
+
55
+ def _require_identity(app_name: str, user_id: str) -> None:
56
+ """Reject empty identities before they reach the database (fail closed)."""
57
+ if not app_name or not str(app_name).strip():
58
+ raise ValueError("app_name must be a non-empty string")
59
+ if not user_id or not str(user_id).strip():
60
+ raise ValueError("user_id must be a non-empty string")
61
+
62
+
63
+ def _parse_memshield_uri(uri: str) -> tuple[str, dict[str, Any]]:
64
+ """Parse a memshield:// or postgresql:// URI into (postgres_dsn, config_overrides)."""
65
+ parsed = urlparse(uri)
66
+ query_params = parse_qs(parsed.query)
67
+ overrides: dict[str, Any] = {}
68
+
69
+ if "tenant_id" in query_params:
70
+ overrides["tenant_id"] = query_params.pop("tenant_id")[0]
71
+ if "agent_clearance" in query_params:
72
+ overrides["agent_clearance"] = int(query_params.pop("agent_clearance")[0])
73
+ if "default_scope" in query_params:
74
+ overrides["default_scope"] = query_params.pop("default_scope")[0]
75
+ if "default_min_clearance" in query_params:
76
+ overrides["default_min_clearance"] = int(query_params.pop("default_min_clearance")[0])
77
+ if "top_k" in query_params:
78
+ overrides["top_k"] = int(query_params.pop("top_k")[0])
79
+ if "table_name" in query_params:
80
+ overrides["table_name"] = query_params.pop("table_name")[0]
81
+ if "simulate_missing_app_filter" in query_params:
82
+ val = query_params.pop("simulate_missing_app_filter")[0].lower()
83
+ overrides["simulate_missing_app_filter"] = val in ("1", "true", "yes")
84
+
85
+ remaining_query = urlencode(query_params, doseq=True)
86
+ dsn = urlunparse(
87
+ (
88
+ "postgresql",
89
+ parsed.netloc,
90
+ parsed.path,
91
+ parsed.params,
92
+ remaining_query,
93
+ parsed.fragment,
94
+ )
95
+ )
96
+ return dsn, overrides
97
+
98
+
99
+ @dataclass
100
+ class QueryTelemetry(Mapping[str, Any]):
101
+ """Telemetry snapshot recorded on `MemShieldMemoryService.last_query_telemetry`.
102
+
103
+ Supports both attribute access (`telemetry.latency_ms`) and dictionary access
104
+ (`telemetry["latency_ms"]`).
105
+ """
106
+
107
+ operation: str
108
+ sql: str
109
+ session_context: dict[str, str]
110
+ latency_ms: float
111
+ row_count: int
112
+ simulate_missing_app_filter: bool
113
+ table_name: str
114
+
115
+ def to_dict(self) -> dict[str, Any]:
116
+ return asdict(self)
117
+
118
+ def __getitem__(self, key: str) -> Any:
119
+ return asdict(self)[key]
120
+
121
+ def __iter__(self) -> Iterator[str]:
122
+ return iter(asdict(self))
123
+
124
+ def __len__(self) -> int:
125
+ return len(asdict(self))
126
+
127
+
128
+ class MemShieldMemoryService(BaseMemoryService):
129
+ """Zero-trust multi-tenant semantic memory service for Google ADK.
130
+
131
+ Enforces tenant, user, and multi-agent clearance boundaries at the PostgreSQL
132
+ kernel level using Row-Level Security (RLS) and transaction-local `set_config`.
133
+
134
+ Blocking work (ONNX embedding and psycopg I/O) runs in a worker thread via
135
+ `asyncio.to_thread`, so the ADK event loop is never stalled by memory calls.
136
+ """
137
+
138
+ def __init__(
139
+ self,
140
+ config: MemShieldConfig | None = None,
141
+ *,
142
+ dsn: str | None = None,
143
+ uri: str | None = None,
144
+ tenant_id: str | None = None,
145
+ agent_clearance: int | None = None,
146
+ default_scope: Literal["private", "tenant_shared"] | None = None,
147
+ default_min_clearance: int | None = None,
148
+ top_k: int | None = None,
149
+ min_similarity: float | None = None,
150
+ embedding_model: str | None = None,
151
+ embedding_dim: int | None = None,
152
+ table_name: str | None = None,
153
+ simulate_missing_app_filter: bool | None = None,
154
+ embed_fn: EmbedFn | None = None,
155
+ pool: ConnectionPool | None = None,
156
+ ) -> None:
157
+ cfg_data: dict[str, Any] = config.model_dump() if config is not None else {}
158
+
159
+ if uri is not None:
160
+ parsed_dsn, uri_overrides = _parse_memshield_uri(uri)
161
+ cfg_data["dsn"] = parsed_dsn
162
+ cfg_data.update(uri_overrides)
163
+
164
+ explicit_overrides = {
165
+ "dsn": dsn,
166
+ "tenant_id": tenant_id,
167
+ "agent_clearance": agent_clearance,
168
+ "default_scope": default_scope,
169
+ "default_min_clearance": default_min_clearance,
170
+ "top_k": top_k,
171
+ "min_similarity": min_similarity,
172
+ "embedding_model": embedding_model,
173
+ "embedding_dim": embedding_dim,
174
+ "table_name": table_name,
175
+ "simulate_missing_app_filter": simulate_missing_app_filter,
176
+ }
177
+ for key, value in explicit_overrides.items():
178
+ if value is not None:
179
+ cfg_data[key] = value
180
+
181
+ self.config = MemShieldConfig(**cfg_data)
182
+ self.embedder = MemShieldEmbedder(
183
+ model_name=self.config.embedding_model,
184
+ expected_dim=self.config.embedding_dim,
185
+ embed_fn=embed_fn,
186
+ )
187
+ self._pool = pool if pool is not None else _get_shared_pool(self.config.dsn)
188
+ self.last_query_telemetry: QueryTelemetry | None = None
189
+
190
+ def _rls_context(self, *, app_name: str, user_id: str) -> dict[str, str]:
191
+ return {
192
+ "app.current_app_name": str(app_name),
193
+ "app.current_tenant_id": str(self.config.tenant_id),
194
+ "app.current_user_id": str(user_id),
195
+ "app.agent_clearance": str(self.config.agent_clearance),
196
+ }
197
+
198
+ @contextmanager
199
+ def rls_transaction(
200
+ self,
201
+ *,
202
+ app_name: str,
203
+ user_id: str,
204
+ ) -> Iterator[psycopg.Cursor[Any]]:
205
+ """Yield a cursor inside an explicit transaction bound to the caller's RLS identity.
206
+
207
+ Identity is bound with `set_config(..., is_local => true)`, so it is discarded when
208
+ the transaction commits or rolls back and never bleeds into the next pooled
209
+ checkout. Any custom SQL executed on the yielded cursor is subject to the RLS
210
+ policies on `memories`, even if that SQL omits its own ownership filters.
211
+ """
212
+ _require_identity(app_name, user_id)
213
+ context = self._rls_context(app_name=app_name, user_id=user_id)
214
+ with self._pool.connection() as conn:
215
+ with conn.transaction():
216
+ with conn.cursor() as cur:
217
+ cur.execute(
218
+ """
219
+ SELECT
220
+ set_config('app.current_app_name', %s, true),
221
+ set_config('app.current_tenant_id', %s, true),
222
+ set_config('app.current_user_id', %s, true),
223
+ set_config('app.agent_clearance', %s, true);
224
+ """,
225
+ (
226
+ context["app.current_app_name"],
227
+ context["app.current_tenant_id"],
228
+ context["app.current_user_id"],
229
+ context["app.agent_clearance"],
230
+ ),
231
+ )
232
+ yield cur
233
+
234
+ async def add_session_to_memory(self, session: Session) -> None:
235
+ """Extract non-empty text parts from `session.events`, embed, and insert with RLS.
236
+
237
+ Idempotent per event: ADK may add the same session several times during its
238
+ lifetime, and previously stored `(session_id, event_id)` pairs are skipped.
239
+ """
240
+ session_meta: dict[str, Any] = {}
241
+ if session.state:
242
+ if "scope" in session.state:
243
+ session_meta["scope"] = session.state["scope"]
244
+ if "min_clearance" in session.state:
245
+ session_meta["min_clearance"] = session.state["min_clearance"]
246
+
247
+ await self.add_events_to_memory(
248
+ app_name=session.app_name,
249
+ user_id=session.user_id,
250
+ events=session.events,
251
+ session_id=session.id,
252
+ custom_metadata=session_meta or None,
253
+ )
254
+
255
+ async def add_events_to_memory(
256
+ self,
257
+ *,
258
+ app_name: str,
259
+ user_id: str,
260
+ events: Sequence[Event],
261
+ session_id: str | None = None,
262
+ custom_metadata: Mapping[str, object] | None = None,
263
+ ) -> None:
264
+ """Append incremental ADK events into RLS-protected semantic memory."""
265
+ base_meta = dict(custom_metadata) if custom_metadata else {}
266
+ entries: list[MemoryEntry] = []
267
+
268
+ for event in events:
269
+ text = _extract_text_from_content(event.content)
270
+ if not text:
271
+ continue
272
+ author = event.author or (
273
+ event.content.role if event.content and event.content.role else "user"
274
+ )
275
+ ev_meta = dict(base_meta)
276
+ if session_id is not None:
277
+ ev_meta.setdefault("session_id", session_id)
278
+ if event.id:
279
+ ev_meta.setdefault("event_id", event.id)
280
+
281
+ entries.append(
282
+ MemoryEntry(
283
+ content=types.Content(
284
+ parts=[types.Part(text=text)],
285
+ role=author,
286
+ ),
287
+ author=author,
288
+ custom_metadata=ev_meta,
289
+ )
290
+ )
291
+
292
+ if not entries:
293
+ return
294
+
295
+ await self.add_memory(
296
+ app_name=app_name,
297
+ user_id=user_id,
298
+ memories=entries,
299
+ custom_metadata=base_meta,
300
+ )
301
+
302
+ async def add_memory(
303
+ self,
304
+ *,
305
+ app_name: str,
306
+ user_id: str,
307
+ memories: Sequence[MemoryEntry],
308
+ custom_metadata: Mapping[str, object] | None = None,
309
+ ) -> None:
310
+ """Directly ingest `MemoryEntry` items into RLS-protected PostgreSQL storage."""
311
+ _require_identity(app_name, user_id)
312
+ await asyncio.to_thread(
313
+ self._add_memory_sync,
314
+ app_name=app_name,
315
+ user_id=user_id,
316
+ memories=memories,
317
+ custom_metadata=custom_metadata,
318
+ )
319
+
320
+ def _add_memory_sync(
321
+ self,
322
+ *,
323
+ app_name: str,
324
+ user_id: str,
325
+ memories: Sequence[MemoryEntry],
326
+ custom_metadata: Mapping[str, object] | None,
327
+ ) -> None:
328
+ base_meta = dict(custom_metadata) if custom_metadata else {}
329
+ prepared: list[tuple[MemoryEntry, str, str, str, int, str | None, dict[str, Any]]] = []
330
+
331
+ for entry in memories:
332
+ text = _extract_text_from_content(entry.content)
333
+ if not text:
334
+ continue
335
+ merged_meta = {**base_meta, **(entry.custom_metadata or {})}
336
+ scope = str(merged_meta.pop("scope", self.config.default_scope))
337
+ min_clearance = int(
338
+ merged_meta.pop("min_clearance", self.config.default_min_clearance)
339
+ )
340
+ session_id = merged_meta.pop("session_id", None)
341
+ author = entry.author or (
342
+ entry.content.role if entry.content and entry.content.role else "user"
343
+ )
344
+ prepared.append(
345
+ (entry, text, author, scope, min_clearance, session_id, merged_meta)
346
+ )
347
+
348
+ if not prepared:
349
+ return
350
+
351
+ texts = [item[1] for item in prepared]
352
+ embeddings = self.embedder.embed_documents(texts)
353
+
354
+ # Ownership columns always come from the authenticated call arguments and the
355
+ # service config, never from caller-supplied metadata.
356
+ table = self.config.table_name
357
+ insert_sql = f"""
358
+ INSERT INTO {table} (
359
+ app_name, tenant_id, user_id, session_id, author,
360
+ scope, min_clearance, content, custom_metadata, embedding
361
+ )
362
+ VALUES (%s, %s, %s, %s, %s, %s, %s, %s, %s::jsonb, %s::vector)
363
+ ON CONFLICT (app_name, tenant_id, user_id, session_id, (custom_metadata->>'event_id'))
364
+ WHERE custom_metadata ? 'event_id'
365
+ DO NOTHING
366
+ RETURNING id
367
+ """
368
+
369
+ t0 = time.perf_counter()
370
+ inserted_count = 0
371
+ with self.rls_transaction(app_name=app_name, user_id=user_id) as cur:
372
+ for (
373
+ entry,
374
+ text,
375
+ author,
376
+ scope,
377
+ min_clearance,
378
+ session_id,
379
+ meta,
380
+ ), vec in zip(prepared, embeddings, strict=True):
381
+ cur.execute(
382
+ insert_sql,
383
+ (
384
+ app_name,
385
+ self.config.tenant_id,
386
+ user_id,
387
+ session_id,
388
+ author,
389
+ scope,
390
+ min_clearance,
391
+ text,
392
+ json.dumps(meta),
393
+ _format_vector(vec),
394
+ ),
395
+ )
396
+ row = cur.fetchone()
397
+ if row is None:
398
+ continue # duplicate (session_id, event_id) already stored
399
+ if entry.id is None:
400
+ entry.id = str(row[0])
401
+ inserted_count += 1
402
+
403
+ latency_ms = (time.perf_counter() - t0) * 1000.0
404
+ self.last_query_telemetry = QueryTelemetry(
405
+ operation="add_memory",
406
+ sql=insert_sql.strip(),
407
+ session_context=self._rls_context(app_name=app_name, user_id=user_id),
408
+ latency_ms=latency_ms,
409
+ row_count=inserted_count,
410
+ simulate_missing_app_filter=self.config.simulate_missing_app_filter,
411
+ table_name=table,
412
+ )
413
+
414
+ async def search_memory(
415
+ self,
416
+ *,
417
+ app_name: str,
418
+ user_id: str,
419
+ query: str,
420
+ ) -> SearchMemoryResponse:
421
+ """Search semantic memories with transaction-local PostgreSQL RLS enforcement."""
422
+ _require_identity(app_name, user_id)
423
+ return await asyncio.to_thread(
424
+ self._search_memory_sync, app_name=app_name, user_id=user_id, query=query
425
+ )
426
+
427
+ def _search_memory_sync(
428
+ self,
429
+ *,
430
+ app_name: str,
431
+ user_id: str,
432
+ query: str,
433
+ ) -> SearchMemoryResponse:
434
+ query_vec = self.embedder.embed_query(query)
435
+ vec_literal = _format_vector(query_vec)
436
+ table = self.config.table_name
437
+
438
+ if self.config.simulate_missing_app_filter:
439
+ sql = f"""
440
+ SELECT
441
+ id, app_name, tenant_id, user_id, session_id, author,
442
+ scope, min_clearance, content, custom_metadata, created_at,
443
+ 1 - (embedding <=> %s::vector) AS similarity
444
+ FROM {table}
445
+ ORDER BY embedding <=> %s::vector
446
+ LIMIT %s
447
+ """
448
+ params: tuple[Any, ...] = (vec_literal, vec_literal, self.config.top_k)
449
+ else:
450
+ # Defense in depth: the application-layer filter mirrors the RLS SELECT
451
+ # policy, so neither layer alone is a single point of failure.
452
+ sql = f"""
453
+ SELECT
454
+ id, app_name, tenant_id, user_id, session_id, author,
455
+ scope, min_clearance, content, custom_metadata, created_at,
456
+ 1 - (embedding <=> %s::vector) AS similarity
457
+ FROM {table}
458
+ WHERE app_name = %s AND tenant_id = %s
459
+ AND (user_id = %s OR (scope = 'tenant_shared' AND min_clearance <= %s))
460
+ ORDER BY embedding <=> %s::vector
461
+ LIMIT %s
462
+ """
463
+ params = (
464
+ vec_literal,
465
+ app_name,
466
+ self.config.tenant_id,
467
+ user_id,
468
+ self.config.agent_clearance,
469
+ vec_literal,
470
+ self.config.top_k,
471
+ )
472
+
473
+ t0 = time.perf_counter()
474
+ with self.rls_transaction(app_name=app_name, user_id=user_id) as cur:
475
+ cur.execute(sql, params)
476
+ rows = cur.fetchall()
477
+ latency_ms = (time.perf_counter() - t0) * 1000.0
478
+
479
+ results: list[MemoryEntry] = []
480
+ for row in rows:
481
+ (
482
+ row_id,
483
+ row_app_name,
484
+ row_tenant_id,
485
+ row_user_id,
486
+ row_session_id,
487
+ row_author,
488
+ row_scope,
489
+ row_min_clearance,
490
+ row_content,
491
+ row_custom_meta,
492
+ row_created_at,
493
+ row_similarity,
494
+ ) = row
495
+
496
+ similarity_val = float(row_similarity)
497
+ if (
498
+ self.config.min_similarity is not None
499
+ and similarity_val < self.config.min_similarity
500
+ ):
501
+ continue
502
+
503
+ meta_dict: dict[str, Any] = (
504
+ dict(row_custom_meta) if isinstance(row_custom_meta, Mapping) else {}
505
+ )
506
+ meta_dict.update(
507
+ {
508
+ "app_name": row_app_name,
509
+ "tenant_id": row_tenant_id,
510
+ "user_id": row_user_id,
511
+ "session_id": row_session_id,
512
+ "scope": row_scope,
513
+ "min_clearance": int(row_min_clearance),
514
+ "similarity": similarity_val,
515
+ }
516
+ )
517
+ role = row_author or "user"
518
+ results.append(
519
+ MemoryEntry(
520
+ id=str(row_id),
521
+ content=types.Content(
522
+ parts=[types.Part(text=str(row_content))],
523
+ role=role,
524
+ ),
525
+ author=role,
526
+ timestamp=row_created_at.isoformat() if row_created_at else None,
527
+ custom_metadata=meta_dict,
528
+ )
529
+ )
530
+
531
+ self.last_query_telemetry = QueryTelemetry(
532
+ operation="search_memory",
533
+ sql=sql.strip(),
534
+ session_context=self._rls_context(app_name=app_name, user_id=user_id),
535
+ latency_ms=latency_ms,
536
+ row_count=len(results),
537
+ simulate_missing_app_filter=self.config.simulate_missing_app_filter,
538
+ table_name=table,
539
+ )
540
+ return SearchMemoryResponse(memories=results)
541
+
542
+ async def delete_memory(
543
+ self,
544
+ *,
545
+ app_name: str,
546
+ user_id: str,
547
+ memory_id: str | None = None,
548
+ ) -> None:
549
+ """Delete a specific memory ID (or all caller-owned memories) under RLS."""
550
+ _require_identity(app_name, user_id)
551
+ await asyncio.to_thread(
552
+ self._delete_memory_sync, app_name=app_name, user_id=user_id, memory_id=memory_id
553
+ )
554
+
555
+ def _delete_memory_sync(
556
+ self,
557
+ *,
558
+ app_name: str,
559
+ user_id: str,
560
+ memory_id: str | None,
561
+ ) -> None:
562
+ table = self.config.table_name
563
+ owner_filter = "app_name = %s AND tenant_id = %s AND user_id = %s"
564
+ owner_params: tuple[Any, ...] = (app_name, self.config.tenant_id, user_id)
565
+ if memory_id is not None:
566
+ sql = f"DELETE FROM {table} WHERE id = %s::uuid AND {owner_filter}"
567
+ params: tuple[Any, ...] = (memory_id, *owner_params)
568
+ else:
569
+ sql = f"DELETE FROM {table} WHERE {owner_filter}"
570
+ params = owner_params
571
+
572
+ t0 = time.perf_counter()
573
+ with self.rls_transaction(app_name=app_name, user_id=user_id) as cur:
574
+ cur.execute(sql, params)
575
+ deleted_count = cur.rowcount if cur.rowcount is not None else 0
576
+ latency_ms = (time.perf_counter() - t0) * 1000.0
577
+
578
+ self.last_query_telemetry = QueryTelemetry(
579
+ operation="delete_memory",
580
+ sql=sql,
581
+ session_context=self._rls_context(app_name=app_name, user_id=user_id),
582
+ latency_ms=latency_ms,
583
+ row_count=deleted_count,
584
+ simulate_missing_app_filter=self.config.simulate_missing_app_filter,
585
+ table_name=table,
586
+ )
587
+
588
+
589
+ def register_adk_uri(scheme: str = "memshield") -> None:
590
+ """Register the `memshield://` URI scheme with Google ADK's ServiceRegistry."""
591
+ from google.adk.cli.service_registry import get_service_registry
592
+
593
+ def _factory(uri: str, **kwargs: Any) -> BaseMemoryService:
594
+ kwargs_copy = dict(kwargs)
595
+ kwargs_copy.pop("agents_dir", None)
596
+ return MemShieldMemoryService(uri=uri, **kwargs_copy)
597
+
598
+ get_service_registry().register_memory_service(scheme, _factory)
@@ -0,0 +1,118 @@
1
+ -- =============================================================================
2
+ -- adk-memshield (v0.1.0) PostgreSQL + pgvector Schema & RLS Policies
3
+ -- =============================================================================
4
+ -- Production schema. Apply this file only; `02_benchmark_baseline.sql` creates an
5
+ -- intentionally unprotected table for benchmarks and must not ship to production.
6
+
7
+ CREATE EXTENSION IF NOT EXISTS vector;
8
+ CREATE EXTENSION IF NOT EXISTS pgcrypto;
9
+
10
+ -- 1. Create non-superuser role to enforce Row-Level Security (RLS)
11
+ -- (PostgreSQL superusers bypass RLS even with FORCE ROW LEVEL SECURITY)
12
+ -- 'memshield_pass' is a local development default: rotate it in any real deployment.
13
+ DO $$
14
+ BEGIN
15
+ IF NOT EXISTS (SELECT FROM pg_roles WHERE rolname = 'memshield_app') THEN
16
+ CREATE ROLE memshield_app LOGIN PASSWORD 'memshield_pass' NOSUPERUSER NOCREATEDB NOCREATEROLE NOBYPASSRLS;
17
+ END IF;
18
+ END
19
+ $$;
20
+
21
+ GRANT USAGE ON SCHEMA public TO memshield_app;
22
+
23
+ -- 2. RLS-Protected Semantic Memory Table
24
+ -- Identity columns reject empty strings: once a pooled connection has used
25
+ -- set_config(..., true), current_setting(..., true) returns '' (not NULL) after the
26
+ -- transaction ends, so an empty-string owner would otherwise match a context-free query.
27
+ CREATE TABLE IF NOT EXISTS memories (
28
+ id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
29
+ app_name TEXT NOT NULL CHECK (app_name <> ''),
30
+ tenant_id TEXT NOT NULL CHECK (tenant_id <> ''),
31
+ user_id TEXT NOT NULL CHECK (user_id <> ''),
32
+ session_id TEXT,
33
+ author TEXT DEFAULT 'user',
34
+ scope TEXT NOT NULL DEFAULT 'private' CHECK (scope IN ('private', 'tenant_shared')),
35
+ min_clearance INTEGER NOT NULL DEFAULT 1 CHECK (min_clearance BETWEEN 1 AND 5),
36
+ content TEXT NOT NULL,
37
+ custom_metadata JSONB NOT NULL DEFAULT '{}'::jsonb,
38
+ embedding vector(384) NOT NULL,
39
+ created_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
40
+ );
41
+
42
+ CREATE INDEX IF NOT EXISTS idx_memories_scope_lookup
43
+ ON memories (app_name, tenant_id, user_id);
44
+
45
+ -- Idempotent session ingestion: ADK may call add_session_to_memory() repeatedly for
46
+ -- the same session, so each (owner, session, event) is stored at most once.
47
+ CREATE UNIQUE INDEX IF NOT EXISTS idx_memories_session_event
48
+ ON memories (app_name, tenant_id, user_id, session_id, (custom_metadata->>'event_id'))
49
+ WHERE custom_metadata ? 'event_id';
50
+
51
+ GRANT SELECT, INSERT, UPDATE, DELETE ON memories TO memshield_app;
52
+
53
+ -- 3. Enable & Force Row-Level Security on `memories`
54
+ ALTER TABLE memories ENABLE ROW LEVEL SECURITY;
55
+ ALTER TABLE memories FORCE ROW LEVEL SECURITY;
56
+
57
+ -- Read: own rows, plus tenant-shared rows at or below the agent's clearance.
58
+ DROP POLICY IF EXISTS memories_isolation_select ON memories;
59
+ CREATE POLICY memories_isolation_select ON memories
60
+ FOR SELECT
61
+ TO memshield_app
62
+ USING (
63
+ app_name = current_setting('app.current_app_name', true)
64
+ AND tenant_id = current_setting('app.current_tenant_id', true)
65
+ AND (
66
+ user_id = current_setting('app.current_user_id', true)
67
+ OR (
68
+ scope = 'tenant_shared'
69
+ AND min_clearance <= COALESCE(NULLIF(current_setting('app.agent_clearance', true), '')::int, 1)
70
+ )
71
+ )
72
+ );
73
+
74
+ -- Write: own rows only. A tenant_shared row may not be labelled above the writing
75
+ -- agent's clearance ("no write-up"), so a Tier-1 agent cannot plant content that
76
+ -- higher-clearance agents would trust as executive memory.
77
+ DROP POLICY IF EXISTS memories_isolation_insert ON memories;
78
+ CREATE POLICY memories_isolation_insert ON memories
79
+ FOR INSERT
80
+ TO memshield_app
81
+ WITH CHECK (
82
+ app_name = current_setting('app.current_app_name', true)
83
+ AND tenant_id = current_setting('app.current_tenant_id', true)
84
+ AND user_id = current_setting('app.current_user_id', true)
85
+ AND (
86
+ scope = 'private'
87
+ OR min_clearance <= COALESCE(NULLIF(current_setting('app.agent_clearance', true), '')::int, 1)
88
+ )
89
+ );
90
+
91
+ DROP POLICY IF EXISTS memories_isolation_update ON memories;
92
+ CREATE POLICY memories_isolation_update ON memories
93
+ FOR UPDATE
94
+ TO memshield_app
95
+ USING (
96
+ app_name = current_setting('app.current_app_name', true)
97
+ AND tenant_id = current_setting('app.current_tenant_id', true)
98
+ AND user_id = current_setting('app.current_user_id', true)
99
+ )
100
+ WITH CHECK (
101
+ app_name = current_setting('app.current_app_name', true)
102
+ AND tenant_id = current_setting('app.current_tenant_id', true)
103
+ AND user_id = current_setting('app.current_user_id', true)
104
+ AND (
105
+ scope = 'private'
106
+ OR min_clearance <= COALESCE(NULLIF(current_setting('app.agent_clearance', true), '')::int, 1)
107
+ )
108
+ );
109
+
110
+ DROP POLICY IF EXISTS memories_isolation_delete ON memories;
111
+ CREATE POLICY memories_isolation_delete ON memories
112
+ FOR DELETE
113
+ TO memshield_app
114
+ USING (
115
+ app_name = current_setting('app.current_app_name', true)
116
+ AND tenant_id = current_setting('app.current_tenant_id', true)
117
+ AND user_id = current_setting('app.current_user_id', true)
118
+ );
@@ -0,0 +1,280 @@
1
+ Metadata-Version: 2.5
2
+ Name: adk-memshield
3
+ Version: 0.1.0
4
+ Summary: Zero-trust multi-tenant semantic memory service for Google ADK backed by PostgreSQL pgvector and Row-Level Security (RLS).
5
+ Project-URL: Homepage, https://github.com/aalishan001/adk-memshield
6
+ Project-URL: Repository, https://github.com/aalishan001/adk-memshield
7
+ Project-URL: Issues, https://github.com/aalishan001/adk-memshield/issues
8
+ Author: Aalishan
9
+ License-Expression: Apache-2.0
10
+ License-File: LICENSE
11
+ Keywords: agents,ai-security,google-adk,memleak,memory,multi-tenant,pgvector,postgresql,row-level-security
12
+ Classifier: Development Status :: 3 - Alpha
13
+ Classifier: Framework :: AsyncIO
14
+ Classifier: Intended Audience :: Developers
15
+ Classifier: Operating System :: OS Independent
16
+ Classifier: Programming Language :: Python :: 3
17
+ Classifier: Programming Language :: Python :: 3 :: Only
18
+ Classifier: Programming Language :: Python :: 3.10
19
+ Classifier: Programming Language :: Python :: 3.11
20
+ Classifier: Programming Language :: Python :: 3.12
21
+ Classifier: Topic :: Database
22
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
23
+ Classifier: Topic :: Security
24
+ Requires-Python: >=3.10
25
+ Requires-Dist: fastembed>=0.3.0
26
+ Requires-Dist: google-adk<3,>=2.12.0
27
+ Requires-Dist: psycopg[binary,pool]>=3.1.0
28
+ Requires-Dist: pydantic>=2.6.0
29
+ Provides-Extra: dev
30
+ Requires-Dist: build>=1.0.0; extra == 'dev'
31
+ Requires-Dist: pytest-asyncio>=0.23.0; extra == 'dev'
32
+ Requires-Dist: pytest>=8.0.0; extra == 'dev'
33
+ Description-Content-Type: text/markdown
34
+
35
+ # `adk-memshield`
36
+
37
+ **Zero-trust multi-tenant semantic memory for Google ADK (`google-adk`), enforced in the database by PostgreSQL `pgvector` + Row-Level Security (RLS).**
38
+
39
+ [![License: Apache-2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](./LICENSE)
40
+ [![Python 3.10+](https://img.shields.io/badge/python-3.10%2B-blue.svg)](./pyproject.toml)
41
+ [![Google ADK](https://img.shields.io/badge/google--adk-%3E%3D2.12-4285F4.svg)](https://github.com/google/adk-python)
42
+
43
+ ---
44
+
45
+ ## Why `adk-memshield`?
46
+
47
+ Multi-tenant and multi-agent `google-adk` deployments often keep long-term semantic memory in one shared vector table, with isolation enforced only by the `WHERE` clause the application remembers to write. That design leaks in two well-documented ways:
48
+
49
+ 1. **Cross-user semantic leakage** — *MemLeak: Cross-User Semantic Leakage in Multi-Tenant AI Agent Memory* ([arXiv:2610.04195](https://arxiv.org/abs/2610.04195)) shows that similarity search over a pooled store surfaces other users' memories without any exploit, and that hard ownership gating is the mitigation that consistently restores the clean baseline.
50
+ 2. **Lateral leakage inside agent swarms** — *AgentLeak* ([arXiv:2602.11510](https://arxiv.org/abs/2602.11510)) shows that internal channels between agents are a major leakage path. `adk-memshield` addresses the memory-retrieval slice of that problem: a low-clearance sub-agent must not retrieve shared tenant memories labelled above its clearance.
51
+
52
+ `adk-memshield` is a drop-in `BaseMemoryService` that moves ownership gating out of application code and into PostgreSQL itself, using **`FORCE ROW LEVEL SECURITY`** and **transaction-local identity (`set_config(..., true)`)**. A retrieval query with a missing or wrong filter returns fewer rows, not someone else's rows.
53
+
54
+ ---
55
+
56
+ ## Architecture
57
+
58
+ ```mermaid
59
+ flowchart LR
60
+ subgraph ADK["Google ADK Runtime"]
61
+ Agent["ADK Agent / Sub-Agent<br/>(app_name, user_id from Runner)"]
62
+ Service["MemShieldMemoryService<br/>(BaseMemoryService)<br/>tenant_id + clearance from config"]
63
+ Embedder["MemShieldEmbedder<br/>(FastEmbed bge-small-en-v1.5, 384d)"]
64
+ end
65
+
66
+ subgraph PG["PostgreSQL 16 + pgvector (role memshield_app, NOBYPASSRLS)"]
67
+ Tx["Explicit transaction"]
68
+ Ctx["Transaction-local identity<br/>set_config('app.*', val, true)"]
69
+ RLS["RLS policies<br/>SELECT / INSERT / UPDATE / DELETE"]
70
+ Vec[("memories<br/>vector(384)")]
71
+ end
72
+
73
+ Agent --> Service
74
+ Service --> Embedder
75
+ Service --> Tx
76
+ Tx --> Ctx
77
+ Ctx --> RLS
78
+ RLS --> Vec
79
+ ```
80
+
81
+ ### Security properties
82
+
83
+ - **No superuser bypass.** Superusers and `BYPASSRLS` roles ignore RLS even under `FORCE ROW LEVEL SECURITY`. [`sql/01_schema.sql`](./sql/01_schema.sql) creates a least-privilege `memshield_app` role (`NOSUPERUSER NOCREATEDB NOCREATEROLE NOBYPASSRLS`); connect with it.
84
+ - **Pool-safe identity.** Every operation runs in an explicit transaction and binds `app.current_app_name`, `app.current_tenant_id`, `app.current_user_id` and `app.agent_clearance` with `set_config(..., true)`. The settings vanish at commit/rollback, so identity never bleeds into the next pooled checkout. All values are bound parameters; the only interpolated identifier (`table_name`) is validated against `^[A-Za-z_][A-Za-z0-9_]*$`.
85
+ - **Fail-closed.** Without identity context a query sees 0 rows. Identity columns carry `CHECK (<col> <> '')` because a pooled connection reports `''` (not `NULL`) for a custom setting once it has been used, and the service rejects empty `app_name`/`user_id` before touching the database.
86
+ - **All four commands are gated.** `SELECT` allows own rows plus `tenant_shared` rows with `min_clearance <= app.agent_clearance`. `INSERT` and `UPDATE` (`USING` + `WITH CHECK`) and `DELETE` (`USING`) are restricted to the caller's own rows, so cross-user poisoning or re-owning a row raises `InsufficientPrivilege`.
87
+ - **No write-up.** An agent cannot write a `tenant_shared` memory labelled above its own clearance, so a Tier-1 agent cannot plant "executive" context that higher-clearance agents would trust.
88
+ - **Defense in depth.** The library's own query also filters by user/scope/clearance, mirroring the policy, so neither layer is a single point of failure. Ownership columns always come from call arguments and config, never from caller-supplied metadata.
89
+ - **Idempotent session ingestion.** ADK may call `add_session_to_memory` repeatedly for one session; a unique index on `(owner, session_id, event_id)` stores each event once.
90
+ - **Non-blocking.** Embedding and database I/O run in a worker thread (`asyncio.to_thread`), so memory calls do not stall the ADK event loop.
91
+
92
+ ---
93
+
94
+ ## Benchmark: unprotected `pgvector` vs. `adk-memshield`
95
+
96
+ [`benchmarks/run_memleak_eval.py`](./benchmarks/run_memleak_eval.py) seeds a synthetic enterprise corpus (**3 tenants, 6 users, 60 memories**: API credentials, compensation, medical/HR records, query-shaped lure memories, and L3/L4 executive memos) into an ordinary table and into the RLS table. Each scenario runs **the same buggy application-layer SQL** against both, inside the same transaction-local identity of the authenticated caller, so the only variable is whether the database enforces ownership. `top_k=3`, 75 queries, 5 repeats.
97
+
98
+ | Scenario | Application-layer mistake | Queries | Unprotected: queries leaking | Unprotected: leaked share of top-k rows | Unprotected: leak at rank 1 | `adk-memshield`: queries leaking | Unprotected clean recall | `adk-memshield` clean recall |
99
+ | :--- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
100
+ | Scenario 1: Shared-Tenant Pool (Incidental Overlap) | Filters `tenant_id` only, omits `user_id`/`scope` | 12 | 100.0% | 44.4% | 8.3% | **0.0%** | 100.0% | **100.0%** |
101
+ | Scenario 2: Adversarial Top-K Planted Lure | Same tenant-only filter; co-tenant plants query-shaped lures | 12 | 100.0% | 58.3% | 58.3% | **0.0%** | 100.0% | **100.0%** |
102
+ | Scenario 3: Tool-Argument `user_id` Override | Correct filter, but `user_id` taken from attacker-controlled tool args | 9 | 100.0% | 81.5% | 100.0% | **0.0%** | n/a | **n/a** |
103
+ | Scenario 4: Low-Clearance Swarm Agent (AgentLeak-style) | Own + `tenant_shared` rows, omits `min_clearance` check | 9 | 100.0% | 44.4% | 66.7% | **0.0%** | 100.0% | **100.0%** |
104
+ | Scenario 5: Total Filter Omission | No `WHERE` clause at all (`simulate_missing_app_filter=True`) | 33 | 100.0% | 74.7% | 69.7% | **0.0%** | 72.7% | **100.0%** |
105
+ | **Mean of Scenarios 1-5** | | 75 | **100.0%** | **60.7%** | **60.6%** | **0.0%** | **93.2%** | **100.0%** |
106
+ | Control: Correct App Filter (no bug) | `MemShieldMemoryService` default filter (user/scope/clearance) | 33 | 0.0% | 0.0% | 0.0% | **0.0%** | 100.0% | **100.0%** |
107
+
108
+ Retrieval latency (database transaction including `set_config`, 60 rows, ms):
109
+
110
+ | Scenario | Unprotected p50 / p95 | `adk-memshield` p50 / p95 |
111
+ | :--- | :---: | :---: |
112
+ | Scenario 1: Shared-Tenant Pool (Incidental Overlap) | 0.74 / 1.19 | 0.78 / 1.38 |
113
+ | Scenario 2: Adversarial Top-K Planted Lure | 0.79 / 1.27 | 0.72 / 1.45 |
114
+ | Scenario 3: Tool-Argument `user_id` Override | 0.83 / 1.35 | 0.84 / 1.42 |
115
+ | Scenario 4: Low-Clearance Swarm Agent (AgentLeak-style) | 0.80 / 1.24 | 0.91 / 1.60 |
116
+ | Scenario 5: Total Filter Omission | 0.72 / 1.22 | 0.73 / 1.22 |
117
+ | Control: Correct App Filter (no bug) | 0.76 / 1.53 | 0.78 / 1.39 |
118
+
119
+ **How to read this**
120
+
121
+ - **Leak rate is saturated by design of fixed top-k.** With no similarity floor, `LIMIT 3` always fills its slots. When a caller has fewer than three relevant rows of their own, a buggy filter fills the remainder with other people's rows, so every buggy query leaked at least one row. The *leaked share* and *rank-1* columns show how much of the retrieved context was unauthorized.
122
+ - **RLS removed every unauthorized row in every scenario**, while the identical buggy SQL leaked on the plain table. The control row confirms the leak classifier reports 0% when the application filter is correct.
123
+ - **Planted lures hijack rank 1, not recall.** Each Scenario 2 lure restates the target question followed by a short malicious payload (a white-box, PoisonedRAG-style attacker who knows the query wording). On the plain table a lure took rank 1 in 58.3% of queries: every roadmap/SLA query, but only 1 of 3 vendor queries and none of the API/KMS queries, where the victim's own note already restates the question almost verbatim. With RLS the lures are never retrieved.
124
+ - **Recall gains are modest and specific.** In Scenarios 1, 2 and 4 the caller's own answer still made the top 3 on the plain table, so RLS changed *what else* was in context (and what ranked first) rather than recall. The recall gain appears under total filter omission (72.7% → 100%), where cross-tenant near-duplicates and lures crowd out the caller's own memories.
125
+ - **Scenario 3 has no authorized answer.** The queries are pure exfiltration attempts, so recall is n/a. RLS turns the attack into an empty or own-rows-only result; it cannot repair the application's wrong `user_id` argument.
126
+ - **Latency.** At 60 rows the RLS overhead is within run-to-run noise. This benchmark does not measure ANN index behaviour at scale (see Limitations).
127
+ - The corpus and queries are synthetic and hand-written. Treat the numbers as a reproducible illustration of failure modes, not as an estimate of real-world leak rates.
128
+
129
+ ---
130
+
131
+ ## Threat model & limitations
132
+
133
+ - **What RLS protects against:** bugs and omissions in application retrieval SQL, prompt-injected tool arguments, and cross-user writes, all *while the service binds the correct identity*.
134
+ - **What it does not protect against:** a compromised application process. `memshield_app` can call `set_config` itself, so code holding the database credentials can claim any identity. Keep identity binding inside this service and keep credentials out of agent-reachable code paths.
135
+ - **Tenant and clearance come from service config, not from the ADK session.** `search_memory` receives only `app_name`/`user_id` from ADK, so run one service instance per tenant/clearance tier (or per Runner).
136
+ - **`tenant_shared` content is trusted by design.** A user may publish a `tenant_shared` memory at or below their own clearance, and every agent in the tenant with sufficient clearance can retrieve it. RLS enforces *who may read*, not whether the content is benign.
137
+ - **ANN indexes and RLS.** The schema ships without a vector index. If you add HNSW/IVFFlat, RLS filters rows *after* the index scan and can return fewer than `top_k` rows. Use pgvector ≥ 0.8 iterative index scans (`SET hnsw.iterative_scan = relaxed_order`) and benchmark recall at your scale.
138
+ - `sql/02_benchmark_baseline.sql` creates an intentionally unprotected table for tests and benchmarks. **Do not apply it in production.**
139
+
140
+ ---
141
+
142
+ ## Installation
143
+
144
+ ```bash
145
+ pip install adk-memshield
146
+ ```
147
+
148
+ From source, for development:
149
+
150
+ ```bash
151
+ pip install -e ".[dev]"
152
+ ```
153
+
154
+ ---
155
+
156
+ ## Quickstart
157
+
158
+ ### 1. Apply the schema
159
+
160
+ Apply [`sql/01_schema.sql`](./sql/01_schema.sql) as a superuser/owner, then rotate the `memshield_app` password. The file also ships inside the wheel:
161
+
162
+ ```python
163
+ from importlib.resources import files
164
+
165
+ schema_sql = (files("adk_memshield") / "sql" / "01_schema.sql").read_text()
166
+ ```
167
+
168
+ For local development, start the bundled Postgres on port `54329`:
169
+
170
+ ```bash
171
+ docker compose -p adk-memshield up -d db
172
+ ```
173
+
174
+ ### 2. Wire it into an ADK `Runner`
175
+
176
+ ```python
177
+ from google.adk.agents.llm_agent import LlmAgent
178
+ from google.adk.runners import Runner
179
+ from google.adk.sessions.in_memory_session_service import InMemorySessionService
180
+ from google.adk.tools.load_memory_tool import load_memory_tool
181
+ from google.adk.tools.preload_memory_tool import preload_memory_tool
182
+
183
+ from adk_memshield import MemShieldConfig, MemShieldMemoryService
184
+
185
+ memory_service = MemShieldMemoryService(
186
+ MemShieldConfig(
187
+ dsn="postgresql://memshield_app:memshield_pass@localhost:54329/memshield",
188
+ tenant_id="tenant_alpha",
189
+ agent_clearance=2,
190
+ )
191
+ )
192
+
193
+ runner = Runner(
194
+ app_name="finance_copilot",
195
+ agent=LlmAgent(
196
+ name="assistant",
197
+ model="gemini-2.5-flash",
198
+ tools=[preload_memory_tool, load_memory_tool],
199
+ ),
200
+ session_service=InMemorySessionService(),
201
+ memory_service=memory_service,
202
+ )
203
+ # ADK's memory tools call memory_service.search_memory(app_name=..., user_id=...) with the
204
+ # Runner's authenticated identity; persist a finished session with:
205
+ # await memory_service.add_session_to_memory(session)
206
+ ```
207
+
208
+ ### 3. Write and search memories directly
209
+
210
+ ```python
211
+ from google.adk.memory.memory_entry import MemoryEntry
212
+ from google.genai import types
213
+
214
+ await memory_service.add_memory(
215
+ app_name="finance_copilot",
216
+ user_id="alice",
217
+ memories=[
218
+ MemoryEntry(
219
+ content=types.Content(parts=[types.Part(text="Wire routing code is SWIFT-ALPHA-99.")]),
220
+ custom_metadata={"scope": "private"},
221
+ ),
222
+ # Visible to tenant agents with clearance >= 2 (the writer's clearance must be >= 2).
223
+ MemoryEntry(
224
+ content=types.Content(parts=[types.Part(text="Treasury approval threshold is $250k.")]),
225
+ custom_metadata={"scope": "tenant_shared", "min_clearance": 2},
226
+ ),
227
+ ],
228
+ )
229
+
230
+ response = await memory_service.search_memory(
231
+ app_name="finance_copilot", user_id="alice", query="What is my wire routing code?"
232
+ )
233
+ for entry in response.memories:
234
+ print(entry.content.parts[0].text, entry.custom_metadata["similarity"])
235
+
236
+ print(memory_service.last_query_telemetry.to_dict()) # SQL, identity context, latency, rows
237
+ ```
238
+
239
+ ### 4. Run custom SQL under the caller's RLS identity
240
+
241
+ ```python
242
+ with memory_service.rls_transaction(app_name="finance_copilot", user_id="alice") as cur:
243
+ cur.execute("SELECT content FROM memories ORDER BY created_at DESC LIMIT 10")
244
+ rows = cur.fetchall() # only rows alice may read, even though the SQL has no WHERE clause
245
+ ```
246
+
247
+ ### 5. `memshield://` URI for the ADK service registry
248
+
249
+ ```python
250
+ from adk_memshield import register_adk_uri
251
+ from google.adk.cli.service_registry import get_service_registry
252
+
253
+ register_adk_uri("memshield")
254
+ memory_service = get_service_registry().create_memory_service(
255
+ "memshield://memshield_app:memshield_pass@localhost:54329/memshield?tenant_id=tenant_alpha&agent_clearance=2"
256
+ )
257
+ ```
258
+
259
+ ---
260
+
261
+ ## Reproducing tests & benchmarks
262
+
263
+ All containers use compose project `adk-memshield` and host port `54329`:
264
+
265
+ ```bash
266
+ # 1. Start pgvector:pg16 (applies sql/01_schema.sql and sql/02_benchmark_baseline.sql on first start)
267
+ docker compose -p adk-memshield up -d db
268
+
269
+ # 2. Run the 8 integration tests, including a real ADK Runner + preload_memory/load_memory test
270
+ docker compose -p adk-memshield run --rm test pytest -v
271
+
272
+ # 3. Run the benchmark
273
+ docker compose -p adk-memshield run --rm test python benchmarks/run_memleak_eval.py
274
+ ```
275
+
276
+ ---
277
+
278
+ ## License
279
+
280
+ Apache License 2.0 — see [`LICENSE`](./LICENSE).
@@ -0,0 +1,9 @@
1
+ adk_memshield/__init__.py,sha256=Plq8LBlAAZWlwX-L9ChANDD9r4aKIudEOTZj_yG0Qlk,426
2
+ adk_memshield/config.py,sha256=FBubNWk0WUJa7w3vwkGRnJOetVpjSP8jz4es_Qvcww4,2788
3
+ adk_memshield/embedder.py,sha256=vitLspPupLGmAoakYHV_OfuNSZm1Fc1Y4nuwyRMFzcg,2244
4
+ adk_memshield/service.py,sha256=yyVFmHl_iM7OS92z12YFjyNU9LNmora965iUPXVp1Qg,22161
5
+ adk_memshield/sql/01_schema.sql,sha256=TLTVxtTqq-SgaQsPMyXaFD4HLzgprQnt2wRdaO3zl_4,5147
6
+ adk_memshield-0.1.0.dist-info/METADATA,sha256=uiSOh56nTHC2BzC3aVHl8cXkuYgCjsk8dv90ceRpMz8,16406
7
+ adk_memshield-0.1.0.dist-info/WHEEL,sha256=W3fkpkm7-wf9vBI5Z-7s0eWkeM-spu78I8Neb98DeEg,87
8
+ adk_memshield-0.1.0.dist-info/licenses/LICENSE,sha256=psuoW8kuDP96RQsdhzwOqi6fyWv0ct8CR6Jr7He_P_k,10173
9
+ adk_memshield-0.1.0.dist-info/RECORD,,
@@ -0,0 +1,4 @@
1
+ Wheel-Version: 1.0
2
+ Generator: hatchling 1.32.4
3
+ Root-Is-Purelib: true
4
+ Tag: py3-none-any
@@ -0,0 +1,176 @@
1
+ Apache License
2
+ Version 2.0, January 2004
3
+ http://www.apache.org/licenses/
4
+
5
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6
+
7
+ 1. Definitions.
8
+
9
+ "License" shall mean the terms and conditions for use, reproduction,
10
+ and distribution as defined by Sections 1 through 9 of this document.
11
+
12
+ "Licensor" shall mean the copyright owner or entity authorized by
13
+ the copyright owner that is granting the License.
14
+
15
+ "Legal Entity" shall mean the union of the acting entity and all
16
+ other entities that control, are controlled by, or are under common
17
+ control with that entity. For the purposes of this definition,
18
+ "control" means (i) the power, direct or indirect, to cause the
19
+ direction or management of such entity, whether by contract or
20
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
21
+ outstanding shares, or (iii) beneficial ownership of such entity.
22
+
23
+ "You" (or "Your") shall mean an individual or Legal Entity
24
+ exercising permissions granted by this License.
25
+
26
+ "Source" form shall mean the preferred form for making modifications,
27
+ including but not limited to software source code, documentation
28
+ source, and configuration files.
29
+
30
+ "Object" form shall mean any form resulting from mechanical
31
+ transformation or translation of a Source form, including but
32
+ not limited to compiled object code, generated documentation,
33
+ and conversions to other media types.
34
+
35
+ "Work" shall mean the work of authorship, whether in Source or
36
+ Object form, made available under the License, as indicated by a
37
+ copyright notice that is included in or attached to the work
38
+ (an example is provided in the Appendix below).
39
+
40
+ "Derivative Works" shall mean any work, whether in Source or Object
41
+ form, that is based on (or derived from) the Work and for which the
42
+ editorial revisions, annotations, elaborations, or other modifications
43
+ represent, as a whole, an original work of authorship. For the purposes
44
+ of this License, Derivative Works shall not include works that remain
45
+ separable from, or merely link (or bind by name) to the interfaces of,
46
+ the Work and Derivative Works thereof.
47
+
48
+ "Contribution" shall mean any work of authorship, including
49
+ the original version of the Work and any modifications or additions
50
+ to that Work or Derivative Works thereof, that is intentionally
51
+ submitted to Licensor for inclusion in the Work by the copyright owner
52
+ or by an individual or Legal Entity authorized to submit on behalf of
53
+ the copyright owner. For the purposes of this definition, "submitted"
54
+ means any form of electronic, verbal, or written communication sent
55
+ to the Licensor or its representatives, including but not limited to
56
+ communication on electronic mailing lists, source code control systems,
57
+ and issue tracking systems that are managed by, or on behalf of, the
58
+ Licensor for the purpose of discussing and improving the Work, but
59
+ excluding communication that is conspicuously marked or otherwise
60
+ designated in writing by the copyright owner as "Not a Contribution."
61
+
62
+ "Contributor" shall mean Licensor and any individual or Legal Entity
63
+ on behalf of whom a Contribution has been received by Licensor and
64
+ subsequently incorporated within the Work.
65
+
66
+ 2. Grant of Copyright License. Subject to the terms and conditions of
67
+ this License, each Contributor hereby grants to You a perpetual,
68
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69
+ copyright license to reproduce, prepare Derivative Works of,
70
+ publicly display, publicly perform, sublicense, and distribute the
71
+ Work and such Derivative Works in Source or Object form.
72
+
73
+ 3. Grant of Patent License. Subject to the terms and conditions of
74
+ this License, each Contributor hereby grants to You a perpetual,
75
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76
+ (except as stated in this section) patent license to make, have made,
77
+ use, offer to sell, sell, import, and otherwise transfer the Work,
78
+ where such license applies only to those patent claims licensable
79
+ by such Contributor that are necessarily infringed by their
80
+ Contribution(s) alone or by combination of their Contribution(s)
81
+ with the Work to which such Contribution(s) was submitted. If You
82
+ institute patent litigation against any entity (including a
83
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
84
+ or a Contribution incorporated within the Work constitutes direct
85
+ or contributory patent infringement, then any patent licenses
86
+ granted to You under this License for that Work shall terminate
87
+ as of the date such litigation is filed.
88
+
89
+ 4. Redistribution. You may reproduce and distribute copies of the
90
+ Work or Derivative Works thereof in any medium, with or without
91
+ modifications, and in Source or Object form, provided that You
92
+ meet the following conditions:
93
+
94
+ (a) You must give any other recipients of the Work or
95
+ Derivative Works a copy of this License; and
96
+
97
+ (b) You must cause any modified files to carry prominent notices
98
+ stating that You changed the files; and
99
+
100
+ (c) You must retain, in the Source form of any Derivative Works
101
+ that You distribute, all copyright, patent, trademark, and
102
+ attribution notices from the Source form of the Work,
103
+ excluding those notices that do not pertain to any part of
104
+ the Derivative Works; and
105
+
106
+ (d) If the Work includes a "NOTICE" text file as part of its
107
+ distribution, then any Derivative Works that You distribute must
108
+ include a readable copy of the attribution notices contained
109
+ within such NOTICE file, excluding those notices that do not
110
+ pertain to any part of the Derivative Works, in at least one
111
+ of the following places: within a NOTICE text file distributed
112
+ as part of the Derivative Works; within the Source form or
113
+ documentation, if provided along with the Derivative Works; or,
114
+ within a display generated by the Derivative Works, if and
115
+ wherever such third-party notices normally appear. The contents
116
+ of the NOTICE file are for informational purposes only and
117
+ do not modify the License. You may add Your own attribution
118
+ notices within Derivative Works that You distribute, alongside
119
+ or as an addendum to the NOTICE text from the Work, provided
120
+ that such additional attribution notices cannot be construed
121
+ as modifying the License.
122
+
123
+ You may add Your own copyright statement to Your modifications and
124
+ may provide additional or different license terms and conditions
125
+ for use, reproduction, or distribution of Your modifications, or
126
+ for any such Derivative Works as a whole, provided Your use,
127
+ reproduction, and distribution of the Work otherwise complies with
128
+ the conditions stated in this License.
129
+
130
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
131
+ any Contribution intentionally submitted for inclusion in the Work
132
+ by You to the Licensor shall be under the terms and conditions of
133
+ this License, without any additional terms or conditions.
134
+ Notwithstanding the above, nothing herein shall supersede or modify
135
+ the terms of any separate license agreement you may have executed
136
+ with Licensor regarding such Contributions.
137
+
138
+ 6. Trademarks. This License does not grant permission to use the trade
139
+ names, trademarks, service marks, or product names of the Licensor,
140
+ except as required for reasonable and customary use in describing the
141
+ origin of the Work and reproducing the content of the NOTICE file.
142
+
143
+ 7. Disclaimer of Warranty. Unless required by applicable law or
144
+ agreed to in writing, Licensor provides the Work (and each
145
+ Contributor provides its Contributions) on an "AS IS" BASIS,
146
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147
+ implied, including, without limitation, any warranties or conditions
148
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149
+ PARTICULAR PURPOSE. You are solely responsible for determining the
150
+ appropriateness of using or redistributing the Work and assume any
151
+ risks associated with Your exercise of permissions under this License.
152
+
153
+ 8. Limitation of Liability. In no event and under no legal theory,
154
+ whether in tort (including negligence), contract, or otherwise,
155
+ unless required by applicable law (such as deliberate and grossly
156
+ negligent acts) or agreed to in writing, shall any Contributor be
157
+ liable to You for damages, including any direct, indirect, special,
158
+ incidental, or consequential damages of any character arising as a
159
+ result of this License or out of the use or inability to use the
160
+ Work (including but not limited to damages for loss of goodwill,
161
+ work stoppage, computer failure or malfunction, or any and all
162
+ other commercial damages or losses), even if such Contributor
163
+ has been advised of the possibility of such damages.
164
+
165
+ 9. Accepting Warranty or Additional Liability. While redistributing
166
+ the Work or Derivative Works thereof, You may choose to offer,
167
+ and charge a fee for, acceptance of support, warranty, indemnity,
168
+ or other liability obligations and/or rights consistent with this
169
+ License. However, in accepting such obligations, You may act only
170
+ on Your own behalf and on Your sole responsibility, not on behalf
171
+ of any other Contributor, and only if You agree to indemnify,
172
+ defend, and hold each Contributor harmless for any liability
173
+ incurred by, or claims asserted against, such Contributor by reason
174
+ of your accepting any such warranty or additional liability.
175
+
176
+ END OF TERMS AND CONDITIONS