Bot Session Learning

Gateway/bot agents learn who the user is across sessions by default — persona, preferences, and insights accrue from plain chat, not just tool-using turns

Any bot or gateway session — Telegram, WhatsApp, Slack, or a custom transport — remembers who each user is across sessions by default, capturing preferences, style, and durable facts from plain conversation, not just from tool-using turns.

Quick Start

1. Zero-config bot

   A bot agent learns from plain chat with no learn wiring — the gateway/bot layer turns on the coordinated posture for you.

   from praisonai_bot.bots import TelegramBot
   from praisonaiagents import Agent
   
   bot = TelegramBot(
       agent=Agent(
           name="Companion",
           instructions="You are a friendly personal assistant.",
       )
   )
   bot.start()

   After a few casual messages, inspect what the bot picked up:

   praisonai memory learn show persona

2. Opt-out (YAML)

   Set learn: false on the channel and the bot behaves exactly as before — no learn stores are written.

   channels:
     telegram:
       token: "${TELEGRAM_TOKEN}"
       learn: false

3. Explicit tuning

   A pre-configured agent always wins — the bot defaults leave it alone.

   from praisonai_bot.bots import TelegramBot
   from praisonaiagents import Agent, LearnConfig
   
   bot = TelegramBot(
       agent=Agent(
           name="Companion",
           instructions="You are a friendly personal assistant.",
           learn=LearnConfig(mode="agentic", nudge_interval=5),
       )
   )
   bot.start()

Coordinated posture

learn=True is one switch that turns on three previously independent mechanisms at once.

You no longer discover and wire three independent knobs — one switch coordinates them.

How It Works

Every user message runs apply_bot_smart_defaults once per agent, then the nudge cadence drives persistence in the background.

• Every user message runs apply_bot_smart_defaults once per agent (on first use).
• After a response, _maybe_emit_nudge increments a per-turn counter; every nudge_interval turns (default 10) it emits a system nudge — regardless of whether tools ran, because nudge_min_tool_iters=0.
• The nudge drives the auto-injected store_learning tool → LearnManager → JSON-backed learn stores under the standard learn path.
• Next turn: get_learn_context() retrieves the accrued persona/insights and auto-injects them into the system prompt.

User Interaction Flow

The bot picks up a stated preference and applies it in a later, fresh session.

What Triggers the Default-On Posture

For a user-supplied memory= instance, the memory-based auto_memory half is skipped so the operator’s chosen backend/intent is preserved; the nudge cadence still fires.

Opt-out

Three equivalent ways disable session learning.

channels:
  telegram:
    token: "${TELEGRAM_TOKEN}"
    learn: false          # or: session_learning: false

from praisonaiagents import Agent

agent = Agent(name="FAQ", instructions="Answer FAQs.", learn=False)

from praisonaiagents.bots.config import BotConfig
from praisonai_bot.bots._defaults import apply_bot_smart_defaults

config = BotConfig()
config.metadata["learn"] = False
apply_bot_smart_defaults(agent, config)

Boolean false, or the strings "false" / "0" / "no" / "off" / "disabled" (case-insensitive) all count as opt-out.

How the Nudge Cadence Changed

nudge_min_tool_iters used to be a hard floor; it is now a tunable override.

• Before: nudge_min_tool_iters=3 was a hard floor — a purely conversational bot never nudged.
• After: <= 0 (the new coordinated default) fires on conversational turns; a positive value keeps the “only after N real tool calls” behaviour for task-oriented agents.

from praisonaiagents import Agent, LearnConfig

Agent(learn=True)                                                # cadence fires every turn
Agent(learn=LearnConfig(mode="agentic", nudge_min_tool_iters=3)) # only after tool work

Safety Properties Preserved

Best Practices

Related