fix: Don't expose the user_id parameter to the LLM

ci/.nim_models_used.json

@@ -11,11 +11,11 @@
     },
     {
       "model": "meta/llama-3.3-70b-instruct",
-      "num_configs": 27
+      "num_configs": 26
     },
     {
       "model": "nvidia/nemotron-3-super-120b-a12b",
-      "num_configs": 16
+      "num_configs": 17
     },
     {
       "model": "meta/llama-3.1-8b-instruct",
@@ -53,7 +53,11 @@
   "embedders": [
     {
       "model": "nvidia/nv-embedqa-e5-v5",
-      "num_configs": 23
+      "num_configs": 22
+    },
+    {
+      "model": "nvidia/nemotron-3-embed-1b",
+      "num_configs": 1
     },
     {
       "model": "nvidia/llama-nemotron-embed-1b-v2",

docs/source/build-workflows/memory.md

@@ -21,6 +21,20 @@ The NeMo Agent Toolkit Memory subsystem is designed to store and retrieve a user
 
 The memory module is designed to be extensible, allowing developers to create custom memory back-ends, providers in NeMo Agent Toolkit terminology.
 
+## User Identity for Memory Tools
+
+The built-in `add_memory`, `get_memory`, and `delete_memory` tools bind every operation to an identity. Set an optional `user_id` in the tool configuration for a fixed service or single-user identity. When
+it is omitted, the tool reads `Context.user_id` for the current invocation. If neither is available, the operation
+fails rather than accessing an unscoped memory namespace.
+
+```yaml
+functions:
+  get_memory:
+    _type: get_memory
+    memory: user_memory
+    user_id: service_user  # Optional; otherwise use the current Context.user_id
+```
+
 ## Included Memory Modules
 The NeMo Agent Toolkit includes four memory module providers, all of which are available as plugins:
 * [Mem0](https://mem0.ai/) which is provided by the [`nvidia-nat-mem0ai`](https://pypi.org/project/nvidia-nat-mem0ai/) plugin.
@@ -32,6 +46,50 @@ The NeMo Agent Toolkit includes four memory module providers, all of which are a
 Additional memory backends are available as community plugins:
 * [Synap](https://maximem.ai) — managed memory layer with user and customer scoping, provided by the [`maximem-synap-nemo-agent-toolkit`](https://pypi.org/project/maximem-synap-nemo-agent-toolkit/) plugin. See `examples/memory/synap/` for usage. ([Open source integration package](https://github.com/maximem-ai/maximem_synap_sdk/tree/main/packages/integrations))
 
+## Authenticating Memory Tool Users
+
+Each of the built-in `add_memory`, `get_memory`, and `delete_memory` tools require configuring an identity source:
+
+- Use `user_id` for a fixed, single-user memory namespace.
+- Use `user_id_resolver` for a multi-user application. Its value is the import path of a trusted, zero-argument Python callable that returns the current authenticated user's stable ID. The callable can be synchronous or asynchronous and is invoked for every memory operation.
+
+For example, application code can obtain a principal that authentication middleware has already verified:
+
+```python
+# my_application/auth.py
+from my_application.request_context import get_authenticated_principal
+
+
+def resolve_memory_user_id() -> str:
+    principal = get_authenticated_principal()
+    if principal is None:
+        raise RuntimeError("An authenticated principal is required")
+    return principal.stable_user_id
+```
+
+Reference that callable from each memory tool:
+
+```yaml
+functions:
+  add_memory:
+    _type: add_memory
+    memory: user_memory
+    user_id_resolver: my_application.auth.resolve_memory_user_id
+  get_memory:
+    _type: get_memory
+    memory: user_memory
+    user_id_resolver: my_application.auth.resolve_memory_user_id
+  delete_memory:
+    _type: delete_memory
+    memory: user_memory
+    user_id_resolver: my_application.auth.resolve_memory_user_id
+```
+
+The resolver is part of the application's trusted computing base. It must derive the ID from authenticated state and
+enforce any required authorization. Do not return an identity supplied by the LLM or copy an unverified request header,
+JWT claim, API key, or cookie value directly. NVIDIA NeMo Agent Toolkit invokes the resolver but does not authenticate
+the value it returns.
+
 ## Automatic Memory Wrapper Agent
 
 The NeMo Agent Toolkit provides an [`auto_memory_agent`](../components/agents/auto-memory-wrapper/index.md) wrapper that adds automatic memory capture and retrieval to any agent without requiring the LLM to invoke memory tools explicitly.
@@ -93,12 +151,17 @@ The automatic memory wrapper agent supports several configuration parameters:
 
 User ID is automatically extracted at runtime for memory isolation via:
 1. `SessionManager.session(user_id=...)` - For production with custom auth middleware (recommended)
-2. `X-User-ID` HTTP header - For testing without middleware
+2. `X-User-ID` HTTP header - Illustrative/testing only; assumes a trusted upstream proxy authenticates the request
+   and injects this header
 3. Console front end `user_id` - Defaults to `"nat_run_user_id"` for `nat run`
 
 Conversation-aware memory backends can also use `conversation_id` to isolate separate conversations for the same user.
 For `nat run`, pass `--conversation_id` when testing independent memory conversations from the CLI.
 
+Never treat a client-supplied `X-User-ID` header as authentication. The header fallback is illustrative only and is
+safe in production only when a trusted upstream proxy removes any client-supplied value and injects the authenticated
+identity.
+
 For detailed configuration and usage examples, refer to the `examples/agents/auto_memory_wrapper/README.md` guide.
 
 ## Examples

docs/source/components/agents/auto-memory-wrapper/auto-memory-wrapper.md

@@ -166,7 +166,8 @@ through the front end or session runtime, not the `auto_memory_agent` workflow b
 ### User ID Extraction Priority
 
 1. **`SessionManager.session(user_id=...)`** - For production with custom auth middleware (recommended)
-2. **`X-User-ID` HTTP header** - For testing without middleware
+2. **`X-User-ID` HTTP header** - Illustrative/testing only; assumes a trusted upstream proxy authenticates the request
+   and injects the header
 3. **Console front end `user_id`** - Defaults to `"nat_run_user_id"` for `nat run`
 
 Conversation-aware memory backends can also use `conversation_id` to isolate separate conversations for the same user.
@@ -198,7 +199,9 @@ async def handle_request(request):
 
 ### Testing: X-User-ID Header
 
-For quick testing without custom middleware:
+This header is not authentication. The following example is illustrative only and assumes that a trusted upstream
+proxy has removed any client-supplied `X-User-ID` value, authenticated the request, and injected the header before
+forwarding it to the application. Do not accept this header directly from untrusted clients.
 
 ```bash
 curl -X POST http://localhost:8000/chat \
@@ -327,7 +330,8 @@ workflow:
 
 ## Important Notes
 
-1. **User ID is runtime/front-end scoped** - Set via `SessionManager.session(user_id=...)`, `X-User-ID`, or `nat run --user_id`
+1. **User ID is runtime/front-end scoped** - Set via `SessionManager.session(user_id=...)` or `nat run --user_id`.
+   `X-User-ID` is illustrative only and requires a trusted upstream proxy to inject it.
 2. **Memory backends are interchangeable** - Works with any implementation of `MemoryEditor` interface
 3. **No memory tools needed** - The wrapped agent does not need explicit memory tools configured
 4. **Transparent to inner agent** - The wrapped agent is unaware of memory operations

docs/source/extend/custom-components/memory.md

@@ -19,6 +19,10 @@ limitations under the License.
 
 This documentation presumes familiarity with the NeMo Agent Toolkit [memory module](../../build-workflows/memory.md), [plugin architecture](../plugins.md), the concept of "function registration" using `@register_function`, and how we define [tool](../../build-workflows/functions-and-function-groups/functions.md#agents-and-tools) and workflow configurations in the NeMo Agent Toolkit config described in the [Creating a New Tool and Workflow](../../get-started/tutorials/create-a-new-workflow.md) tutorial.
 
+For applications that expose the built-in memory tools to multiple authenticated users, see
+[Authenticating Memory Tool Users](../../build-workflows/memory.md#authenticating-memory-tool-users). Configure a trusted
+`user_id_resolver`.
+
 ## Key Memory Module Components
 
 * **Memory Data Models**
@@ -190,12 +194,14 @@ functions:
   add_memory:
     _type: add_memory
     memory: saas_memory
+    user_id: user_12
     description: |
       Add any facts about user preferences to long term memory. Always use this if users mention a preference.
       The input to this tool should be a string that describes the user's preference, not the question or answer.
   get_memory:
     _type: get_memory
     memory: saas_memory
+    user_id: user_12
     description: |
       Always call this tool before calling any other tools, even if the user does not mention to use it.
       The question should be about user preferences which will help you format your response.
@@ -214,6 +220,8 @@ Explanation:
 
 - We define a memory entry named `saas_memory` with `_type: mem0_memory`, using the [Mem0](https://mem0.ai/) provider included in the [`nvidia-nat-mem0ai`](https://pypi.org/project/nvidia-nat-mem0ai/) plugin.
 - Then we define two tools (functions in NeMo Agent Toolkit terminology) that reference `saas_memory`: `add_memory` and `get_memory`.
+- The optional `user_id` is a fixed identity for these tools. If it is omitted, the tools use the current invocation's
+  `Context.user_id` and fail if no identity is available. The LLM does not provide this value.
 - Finally, the `agent_memory` workflow references these two tool names.
 
 ### Automatic Memory with the Auto-Memory Wrapper

examples/RAG/simple_rag/configs/milvus_memory_rag_config.yml

@@ -44,12 +44,14 @@ functions:
   add_memory:
     _type: add_memory
     memory: saas_memory
+    user_id: user_12
     description: |
       Add any facts about user preferences to long term memory. Always use this if users mention a preference.
       The input to this tool should be a string that describes the user's preference, not the question or answer.
   get_memory:
     _type: get_memory
     memory: saas_memory
+    user_id: user_12
     description: |
       Always call this tool before calling any other tools, even if the user does not mention to use it.
       The question should be about user preferences which will help you format your response.
@@ -86,16 +88,14 @@ workflow:
 
     IMPORTANT MEMORY TOOL REQUIREMENTS:
     1. You MUST call get_memory tool FIRST, before calling any other tools
-    2. You MUST use user_id "user_12" for all memory operations
-    3. You MUST include ALL required parameters when calling memory tools
-    4. When calling add_memory or get_memory, you MUST use the exact format as below, don't include any other content,
+    2. You MUST include ALL required parameters when calling memory tools
+    3. When calling add_memory or get_memory, you MUST use the exact format as below, don't include any other content,
       and make sure the input is a valid JSON object.
 
     For get_memory tool, you MUST use this exact format:
     {{
         "query": "user preferences",
-        "top_k": 1,
-        "user_id": "user_12"
+        "top_k": 1
     }}
 
     For add_memory tool, you MUST use this exact format:
@@ -110,7 +110,6 @@ workflow:
                 "content": "Hello Alex! I've noted you are looking for a trip to New York."
             }}
         ],
-        "user_id": "user_12",
         "metadata": {{
             "key_value_pairs": {{
                 "type": "travel",

examples/RAG/simple_rag/configs/milvus_memory_rag_tools_config.yml

@@ -44,12 +44,14 @@ functions:
   add_memory:
     _type: add_memory
     memory: saas_memory
+    user_id: user_12
     description: |
       Add any facts about user preferences to long term memory. Always use this if users mention a preference.
       The input to this tool should be a string that describes the user's preference, not the question or answer.
   get_memory:
     _type: get_memory
     memory: saas_memory
+    user_id: user_12
     description: |
       Always call this tool before calling any other tools, even if the user does not mention to use it.
       The question should be about user preferences which will help you format your response.

examples/agents/auto_memory_wrapper/README.md

@@ -133,7 +133,8 @@ User ID is extracted at runtime for memory isolation. Configure it through the f
 ### User ID Extraction Priority
 
 1. **`SessionManager.session(user_id=...)`** - For production with custom auth middleware (recommended)
-2. **`X-User-ID` HTTP header** - For testing without middleware
+2. **`X-User-ID` HTTP header** - Illustrative/testing only; assumes a trusted upstream proxy authenticates the request
+   and injects the header
 3. **Console front end `user_id`** - Defaults to `"nat_run_user_id"` for `nat run`
 
 Conversation-aware memory backends can also use `conversation_id` to isolate separate conversations for the same user.
@@ -163,9 +164,11 @@ async def handle_request(request):
     return result
 ```
 
-### Testing: X-User-ID Header
+### Illustrative X-User-ID Header
 
-For quick testing without custom middleware:
+This header is not authentication. The following example is illustrative only and assumes that a trusted upstream
+proxy has removed any client-supplied `X-User-ID` value, authenticated the request, and injected the header before
+forwarding it to the application. Do not accept this header directly from untrusted clients.
 
 ```bash
 curl -X POST http://localhost:8000/chat \
@@ -213,7 +216,8 @@ workflow:
 
 ## Important Notes
 
-1. **User ID is runtime/front-end scoped** - Set via `SessionManager.session(user_id=...)`, `X-User-ID`, or `nat run --user_id`
+1. **User ID is runtime/front-end scoped** - Set via `SessionManager.session(user_id=...)` or `nat run --user_id`.
+   `X-User-ID` is illustrative only and requires a trusted upstream proxy to inject it.
 2. **Memory backends are interchangeable** - Works with any implementation of `MemoryEditor` interface
 
 ## Examples

examples/frameworks/semantic_kernel_demo/src/nat_semantic_kernel_demo/configs/config.yml

@@ -34,13 +34,15 @@ functions:
   add_memory:
     _type: add_memory
     memory: saas_memory
+    user_id: user_1
     description: |
       Add any facts about user preferences to long term memory. Always use this if users mention a preference.
       The input to this tool should be a string that describes the user's preference, not the question or answer.
 
   get_memory:
     _type: get_memory
     memory: saas_memory
+    user_id: user_1
     description: |
       Always call this tool before calling any other tools, even if the user does not mention to use it.
       The question should be about user preferences which will help you format your response.
@@ -82,16 +84,14 @@ workflow:
     You have access to long term memory.
     IMPORTANT MEMORY TOOL REQUIREMENTS:
       1. You MUST call get_memory tool FIRST, before calling any other tools
-      2. You MUST use user_id "user_1" for all memory operations
-      3. You MUST include ALL required parameters when calling memory tools
-      4. When calling add_memory or get_memory, you MUST use the exact format as below, don't include any other content,
+      2. You MUST include ALL required parameters when calling memory tools
+      3. When calling add_memory or get_memory, you MUST use the exact format as below, don't include any other content,
         and make sure the input is a valid JSON object.
 
     For get_memory tool, you MUST use this exact format:
     {
         "query": "user preferences",
-        "top_k": 1,
-        "user_id": "user_1"
+        "top_k": 1
     }
 
     For add_memory tool, you MUST use this exact format:
@@ -106,7 +106,6 @@ workflow:
                 "content": "Hello Alex! I've noted you are looking for a trip to New York."
             }
         ],
-        "user_id": "user_1",
         "metadata": {
             "key_value_pairs": {
                 "type": "travel",

examples/memory/memmachine/memmachine_memory_example.ipynb

@@ -435,17 +435,19 @@
     "  get_memory:\n",
     "    _type: get_memory\n",
     "    memory: memmachine_memory\n",
+    "    user_id: \"{user_id}\"\n",
     "    description: |\n",
     "      Retrieve memories relevant to a query. Always call this tool first to check\n",
     "      for existing user preferences or facts.\n",
-    "      Use the exact JSON format with user_id, query, and top_k parameters.\n",
+    "      Use the exact JSON format with query and top_k parameters.\n",
     "\n",
     "  add_memory:\n",
     "    _type: add_memory\n",
     "    memory: memmachine_memory\n",
+    "    user_id: \"{user_id}\"\n",
     "    description: |\n",
     "      Add facts about user preferences or information to long-term memory.\n",
-    "      Use the exact JSON format with user_id, memory, conversation (optional), metadata, and tags.\n",
+    "      Use the exact JSON format with memory, conversation (optional), metadata, and tags.\n",
     "\n",
     "workflow:\n",
     "  _type: react_agent\n",