how to mange state in agent

Thanks for reaching out, arpan. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

run time state mangment for storterm memory track

Thanks — I’ll help with runtime short-term memory tracking. Quick clarification: do you want a Python or TypeScript example for LiveKit AgentSession

userdata

, and is the memory per-user session or shared across sessions? If you confirm, I’ll give a concise example and recommended patterns (initialization, read/update, retention/expiry, and when to persist to long-term storage). , with interactive elements

python

Thanks — I'll show a concise Python pattern for short-term (runtime) memory using

AgentSession.userdata

. 1) Define a typed

userdata

dataclass for short-term memory:

from dataclasses import dataclass, field
from typing import List
import time

@dataclass
class ShortTermMemory:
    recent_messages: List[str] = field(default_factory=list)
    last_updated: float = field(default_factory=lambda: time.time())
    ttl_seconds: int = 300  # keep recent memory for 5 minutes

2) Initialize the AgentSession with `userdata`:

session = AgentSession(userdata=ShortTermMemory())

3) Access, update, and enforce TTL inside handlers/tools:

def add_to_memory(session, message: str):
    mem: ShortTermMemory = session.userdata
    # drop expired memory
    if time.time() - mem.last_updated > mem.ttl_seconds:
        mem.recent_messages.clear()
    # append and trim (keep small window)
    mem.recent_messages.append(message)
    if len(mem.recent_messages) > 10:
        mem.recent_messages.pop(0)
    mem.last_updated = time.time()
    session.userdata = mem  # persist back to session

4) Read memory for decisions:

def get_recent_context(session) -> str:
    mem: ShortTermMemory = session.userdata
    return " ".join(mem.recent_messages)

Best practices (brief): • Keep

userdata

small and focused (recent messages, last intent, short flags). • Use a TTL and bounded length to avoid unbounded memory growth. • Treat

userdata

as the single source of runtime session state; agents remain stateless beyond it. • Persist to long-term storage only when you need history beyond short-term (e.g., conversation logs). If you want, I can provide a full handler example wired into your AgentSession lifecycle or adjust TTL/size recommendations for your expected traffic. Which would you prefer? , with interactive elements