Subagent memory modes

Sub-agent spawning has two names because it has two different model-memory contracts: amnemon is tabula rasa; mnemon remembers the parent context and appends a fresh prompt at the end.

Both tools keep the same operational skeleton:

• a child Agent is born under the spawning agent’s parent edge;
• the shell is still a value-snapshot via Shell::fork_session;
• authority is still parent ⊓ resolve_base(permissions);
• the child runs detached through the same drive loop and returns later through the parent’s inbox.

The distinction is only the child’s provider transcript:

• amnemon starts with no parent conversation. It receives only the launch prompt as its first user turn. Use it when blank context is a feature.
• mnemon imports the parent’s model-visible context, reuses the parent’s current provider selection, and appends the tool call’s prompt as the fresh final user prompt. Use it when the child should stand inside the current conversation and hit provider prompt caches.

Because a tool call is dispatched while the parent may be waiting for that very tool result, mnemon must not inherit a dangling assistant tool-call frame. AgentLog::inherited_context_messages drops the unanswered assistant tool-call event when the parent is in AwaitingToolResults, so the child forks the request context rather than an invalid protocol prefix.

This names the two costs honestly. A remembered sub-agent spends less rebuilding context but carries the parent’s assumptions; a blank sub-agent pays the full context boot but gives a genuine fresh read.

Links

agents, agent, tools, provider, async-agent-tool.