What LangMem asks the model

A memory manager is a prompt and a set of tools. Printing the request shows the conversation wrapped in a session tag and a Memory tool the model must call.

This subclass prints every request before answering it:

class ShowRequest(MemoryModel):
    def _generate(self, messages, stop=None, run_manager=None, tools=None, **kwargs):
        print("tools:", [tool["function"]["name"] for tool in tools])
        for message in messages:
            print(f"--- {message.type}")
            print(message.content)
        return super()._generate(messages, stop, run_manager, tools, **kwargs)

manager = create_memory_manager(ShowRequest(), instructions="Keep facts that help support this customer.")
manager.invoke({"messages": [{"role": "user", "content": "Please email me."}]})

• One tool, Memory, whose arguments are the fields of the Memory model: "call this tool once for each new memory".
• A short system message.
• A human message with your instructions, a paragraph telling the model to enrich and prune memories in one set of tool calls, then the conversation inside a <session_...> tag.

Without instructions, LangMem uses a long default prompt about semantic, procedural and episodic memory. Your own instructions replace it, and are the main way to say what is worth remembering for your app.

With existing memories

manager = create_memory_manager(ShowRequest())
manager.invoke({"messages": [{"role": "user", "content": "Text me instead."}], "existing": [("m-1", Memory(content="Wants us to email them"))]})

Passing existing changes the request: the system message lists each memory in an <instance id=...> block, and a PatchDoc tool appears for changing them with JSON Patch operations. Lesson 7 is about that.