Skip to content

Agent Initialization

The Engine runs agent initialization before Game Master and simulation initialization. This phase prepares agent-owned state such as configured raw memories or formative memories, then calls each agent's initialize(...) hook with a typed AgentInitializationContext.

There is no initializer Game Master and no GM memory bank in the native runtime.

Built-Ins

Built-in Config Behavior
default / raw_memory sim.initialization.agents.built_in: raw_memory Inject configured shared and per-agent memories.
formative_memory sim.initialization.agents.built_in: formative_memory Generate per-agent formative memories with the configured model, then inject them.
none sim.initialization.agents.built_in: none Skip agent initialization. 
sim:
  initialization:
    agents:
      built_in: raw_memory
      class_path: null
      params: {}

Custom Initializers

Subclass AgentInitializer and implement initialize(...):

from silisocs.initialization.agents import AgentInitializer


class MyAgentInitializer(AgentInitializer):
    def initialize(self, *, agents, model, context):
        for agent in agents:
            agent.initialize({"memories": [f"Welcome, {agent.name}."]})

Contract details:

  • input agents: live native or adapter-wrapped agents;
  • input model: initializer model, used only when the initializer needs LLM generation;
  • input context: AgentInitializationContext with shared memories, per-agent memories, initial observations, roles, bios, and world data;
  • output: None; mutate agents by calling agent.initialize(...) or agent.observe(...).

Do not create backend users, social graphs, or seed posts here. Backend setup belongs to the Game Master's initialize component; seed posts belong to simulation initialization.

Then configure it:

sim:
  initialization:
    agents:
      class_path: my_package.my_module.MyAgentInitializer
      params: {}