Skip to content

Conversation State

Runlet provides state store primitives, but it does not impose a built-in conversation memory framework.

That boundary is intentional.

What Runlet provides

Runlet includes:

  • StateScope
  • StateStore
  • InMemoryStateStore

These are low-level storage primitives. They are useful for building application-specific session behavior.

What applications should own

Applications should decide:

  • how to persist message history
  • how many prior turns to keep
  • when to summarize old turns
  • whether tool results become durable memory
  • how to partition state by user, thread, or session

Use a session helper that:

  1. loads prior messages from a state store
  2. appends the new user message
  3. calls Runtime.run_request() or Runtime.stream_request()
  4. saves the updated conversation history

Example shape

from runlet import InMemoryStateStore, Message, StateScope
from runlet.core.models import ModelRequest


store = InMemoryStateStore()
scope = StateScope(kind="chat_session", key="session-1")

state = await store.load(scope)
history = state.get("messages", [])

messages = [
    *history,
    Message.user("那我刚才说了什么？"),
]

request = ModelRequest(messages=messages)
result = await runtime.run_request(agent, request)

Why this stays out of Runtime

If Runtime owned conversation memory directly, the core would quickly absorb:

  • session semantics
  • message retention policy
  • summarization policy
  • retrieval policy
  • user-level persistence policy

That would push Runlet toward a larger application framework, which is outside its intended scope.

Next step

See Runtime Boundary.

For a runnable end-to-end example, see Full Example.