Skip to main content

Local Runtime

What the Local Runtime Does​

The default LocalRuntime wires up core services and automates common tasks:

  • Registers agents from their directories and manages IDs
  • Routes messages and builds a minimal AgentContext
  • Automatically stores incoming messages to conversation history
  • Attaches recent conversation_history to the context
  • Provides a history_provider for lazy history loading
  • Lazily initializes a memory manager and model manager

Automatic Conversation Storage​

On each route_message, the runtime:

mm = self.get_memory_manager()
await mm.store_conversation(ctx.conversation_id, message)
ctx.conversation_history = await mm.get_conversation(ctx.conversation_id, limit=50)
from l6e_forge.memory.conversation.provider import ConversationHistoryProvider
ctx.history_provider = ConversationHistoryProvider(mm)

Agents can immediately use context.conversation_history or call the provider for more.

Memory Manager Selection​

The runtime picks defaults then upgrades if providers are available:

  • Vector store: In-memory by default; Qdrant if QDRANT_URL or AF_MEMORY_PROVIDER=qdrant. Backends may support multiple collections.
  • Embeddings: Prefer Ollama (OLLAMA_HOST), then LM Studio (LMSTUDIO_HOST), else a mock embedder.
  • Conversation store: Postgres if AF_DB_URL is set; otherwise in-memory.

Model Manager Selection​

The model manager is created from endpoints loaded via workspace config (forge.toml). Override the default with AF_DEFAULT_PROVIDER.

Environment Variables​

  • QDRANT_URL, AF_MEMORY_PROVIDER
  • OLLAMA_HOST, LMSTUDIO_HOST
  • AF_DB_URL
  • AF_DEFAULT_PROVIDER

Practical Usage in an Agent​

Because the runtime stores conversation messages automatically and attaches recent history, agents can focus on logic:

async def handle_message(self, message, context):
    # History is already available
    recent = context.conversation_history
    mm = self.runtime.get_memory_manager()
    # Store/query vector memory as needed
    await mm.store_vector(namespace=self.name, key=str(message.message_id), content=message.content)
    hits = await mm.search_vectors(namespace=self.name, query=message.content, limit=5)
    # Build a reply using recall + history
    recall = "\n".join(f"- {m.content}" for m in hits)
    return AgentResponse(content=f"You said: {message.content}\n\n{recall}", agent_id=self.name, response_time=0.0)

This behavior is designed for a great local-first developer experience while remaining pluggable via adapters and environment variables.