chore: Merge OSS (#512)

letta/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.6.6"
+__version__ = "0.6.7"
 
 # import clients
 from letta.client.client import LocalClient, RESTClient, create_client

letta/agent.py

@@ -224,8 +224,8 @@ def execute_tool_and_persist_state(self, function_name: str, function_args: dict
                 )
                 function_response, updated_agent_state = sandbox_run_result.func_return, sandbox_run_result.agent_state
                 assert orig_memory_str == self.agent_state.memory.compile(), "Memory should not be modified in a sandbox tool"
-
-                self.update_memory_if_change(updated_agent_state.memory)
+                if updated_agent_state is not None:
+                    self.update_memory_if_change(updated_agent_state.memory)
         except Exception as e:
             # Need to catch error here, or else trunction wont happen
             # TODO: modify to function execution error
@@ -238,7 +238,7 @@ def execute_tool_and_persist_state(self, function_name: str, function_args: dict
     def _get_ai_reply(
         self,
         message_sequence: List[Message],
-        function_call: str = "auto",
+        function_call: Optional[str] = None,
         first_message: bool = False,
         stream: bool = False,  # TODO move to config?
         empty_response_retry_limit: int = 3,
@@ -1029,6 +1029,7 @@ def get_context_window(self) -> ContextWindowOverview:
             num_archival_memory=agent_manager_passage_size,
             num_recall_memory=message_manager_size,
             num_tokens_external_memory_summary=num_tokens_external_memory_summary,
+            external_memory_summary=external_memory_summary,
             # top-level information
             context_window_size_max=self.agent_state.llm_config.context_window,
             context_window_size_current=num_tokens_used_total,

letta/cli/cli_config.py

@@ -60,7 +60,6 @@ def list(arg: Annotated[ListChoice, typer.Argument]):
         table.field_names = ["Name", "Text"]
         for human in client.list_humans():
             table.add_row([human.template_name, human.value.replace("\n", "")[:100]])
-        print(table)
     elif arg == ListChoice.personas:
         """List all personas"""
         table.field_names = ["Name", "Text"]

letta/llm_api/helpers.py

@@ -250,6 +250,8 @@ def unpack_all_inner_thoughts_from_kwargs(
 
 def unpack_inner_thoughts_from_kwargs(choice: Choice, inner_thoughts_key: str) -> Choice:
     message = choice.message
+    rewritten_choice = choice  # inner thoughts unpacked out of the function
+
     if message.role == "assistant" and message.tool_calls and len(message.tool_calls) >= 1:
         if len(message.tool_calls) > 1:
             warnings.warn(f"Unpacking inner thoughts from more than one tool call ({len(message.tool_calls)}) is not supported")
@@ -271,14 +273,18 @@ def unpack_inner_thoughts_from_kwargs(choice: Choice, inner_thoughts_key: str) -
                     warnings.warn(f"Overwriting existing inner monologue ({new_choice.message.content}) with kwarg ({inner_thoughts})")
                 new_choice.message.content = inner_thoughts
 
-                return new_choice
+                # update the choice object
+                rewritten_choice = new_choice
             else:
                 warnings.warn(f"Did not find inner thoughts in tool call: {str(tool_call)}")
-                return choice
 
         except json.JSONDecodeError as e:
             warnings.warn(f"Failed to strip inner thoughts from kwargs: {e}")
             raise e
+    else:
+        warnings.warn(f"Did not find tool call in message: {str(message)}")
+
+    return rewritten_choice
 
 
 def is_context_overflow_error(exception: Union[requests.exceptions.RequestException, Exception]) -> bool:

letta/llm_api/llm_api_tools.py

@@ -94,7 +94,7 @@ def create(
     user_id: Optional[str] = None,  # option UUID to associate request with
     functions: Optional[list] = None,
     functions_python: Optional[dict] = None,
-    function_call: str = "auto",
+    function_call: Optional[str] = None,  # see: https://platform.openai.com/docs/api-reference/chat/create#chat-create-tool_choice
     # hint
     first_message: bool = False,
     force_tool_call: Optional[str] = None,  # Force a specific tool to be called
@@ -132,10 +132,19 @@ def create(
 
     # openai
     if llm_config.model_endpoint_type == "openai":
+
         if model_settings.openai_api_key is None and llm_config.model_endpoint == "https://api.openai.com/v1":
             # only is a problem if we are *not* using an openai proxy
             raise LettaConfigurationError(message="OpenAI key is missing from letta config file", missing_fields=["openai_api_key"])
 
+        if function_call is None and functions is not None and len(functions) > 0:
+            # force function calling for reliability, see https://platform.openai.com/docs/api-reference/chat/create#chat-create-tool_choice
+            # TODO(matt) move into LLMConfig
+            if llm_config.model_endpoint == "https://inference.memgpt.ai":
+                function_call = "auto"  # TODO change to "required" once proxy supports it
+            else:
+                function_call = "required"
+
         data = build_openai_chat_completions_request(llm_config, messages, user_id, functions, function_call, use_tool_naming, max_tokens)
         if stream:  # Client requested token streaming
             data.stream = True

letta/schemas/letta_response.py

@@ -66,7 +66,7 @@ def get_formatted_content(msg):
                 return f'<div class="content"><span class="function-name">{html.escape(msg.function_call.name)}</span>({args})</div>'
             elif msg.message_type == "tool_call_message":
                 args = format_json(msg.tool_call.arguments)
-                return f'<div class="content"><span class="function-name">{html.escape(msg.function_call.name)}</span>({args})</div>'
+                return f'<div class="content"><span class="function-name">{html.escape(msg.tool_call.name)}</span>({args})</div>'
             elif msg.message_type == "function_return":
                 return_value = format_json(msg.function_return)
                 # return f'<div class="status-line">Status: {html.escape(msg.status)}</div><div class="content">{return_value}</div>'

letta/schemas/memory.py

@@ -30,6 +30,9 @@ class ContextWindowOverview(BaseModel):
     num_tokens_external_memory_summary: int = Field(
         ..., description="The number of tokens in the external memory summary (archival + recall metadata)."
     )
+    external_memory_summary: str = Field(
+        ..., description="The metadata summary of the external memory sources (archival + recall metadata)."
+    )
 
     # context window breakdown (in tokens)
     # this should all add up to context_window_size_current

letta/services/agent_manager.py

@@ -388,7 +388,7 @@ def rebuild_system_prompt(self, agent_id: str, actor: PydanticUser, force=False,
         curr_memory_str = agent_state.memory.compile()
         if curr_memory_str in curr_system_message_openai["content"] and not force:
             # NOTE: could this cause issues if a block is removed? (substring match would still work)
-            logger.info(
+            logger.debug(
                 f"Memory hasn't changed for agent id={agent_id} and actor=({actor.id}, {actor.name}), skipping system prompt rebuild"
             )
             return agent_state

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "letta"
-version = "0.6.6"
+version = "0.6.7"
 packages = [
     {include = "letta"}
 ]

tests/helpers/endpoints_helper.py

@@ -117,7 +117,11 @@ def check_first_response_is_valid_for_llm_endpoint(filename: str) -> ChatComplet
     choice = response.choices[0]
 
     # Ensure that the first message returns a "send_message"
-    validator_func = lambda function_call: function_call.name == "send_message" or function_call.name == "archival_memory_search"
+    validator_func = (
+        lambda function_call: function_call.name == "send_message"
+        or function_call.name == "archival_memory_search"
+        or function_call.name == "core_memory_append"
+    )
     assert_contains_valid_function_call(choice.message, validator_func)
 
     # Assert that the message has an inner monologue

tests/integration_test_agent_tool_graph.py

@@ -38,7 +38,7 @@ def second_secret_word(prev_secret_word: str):
         prev_secret_word (str): The secret word retrieved from calling first_secret_word.
     """
     if prev_secret_word != "v0iq020i0g":
-        raise RuntimeError(f"Expected secret {"v0iq020i0g"}, got {prev_secret_word}")
+        raise RuntimeError(f"Expected secret {'v0iq020i0g'}, got {prev_secret_word}")
 
     return "4rwp2b4gxq"
 
@@ -51,7 +51,7 @@ def third_secret_word(prev_secret_word: str):
         prev_secret_word (str): The secret word retrieved from calling second_secret_word.
     """
     if prev_secret_word != "4rwp2b4gxq":
-        raise RuntimeError(f"Expected secret {"4rwp2b4gxq"}, got {prev_secret_word}")
+        raise RuntimeError(f'Expected secret "4rwp2b4gxq", got {prev_secret_word}')
 
     return "hj2hwibbqm"
 
@@ -64,7 +64,7 @@ def fourth_secret_word(prev_secret_word: str):
         prev_secret_word (str): The secret word retrieved from calling third_secret_word.
     """
     if prev_secret_word != "hj2hwibbqm":
-        raise RuntimeError(f"Expected secret {"hj2hwibbqm"}, got {prev_secret_word}")
+        raise RuntimeError(f"Expected secret {'hj2hwibbqm'}, got {prev_secret_word}")
 
     return "banana"
 

tests/test_server.py

@@ -508,6 +508,7 @@ def test_get_context_window_overview(server: SyncServer, user, agent_id):
     assert overview.num_archival_memory is not None
     assert overview.num_recall_memory is not None
     assert overview.num_tokens_external_memory_summary is not None
+    assert overview.external_memory_summary is not None
     assert overview.num_tokens_system is not None
     assert overview.system_prompt is not None
     assert overview.num_tokens_core_memory is not None