Flowstate

Flowstate is a small Python library for maintaining behavioral context for AI agents.

Instead of retrieval context from documents or knowledge context from static reference data, Flowstate keeps the user's live workflow state structured and current. Your agent reads it as background context before responding.

Install

pip install flowstate-ai

Quick start

Create one context per session and update it as the user moves through your product:

from flowstate import Context

ctx = Context(scope={"user_id": "traveller-301"}, max_steps=4)

# traveller opens their booking and sees a delay
ctx.set_fact("flight", "BA 442 — London Heathrow to Amsterdam")
ctx.set_fact("scheduled_departure", "16:45")
ctx.set_fact("status", "Delayed — 2 hours (new departure: 18:45)")
ctx.set_fact("gate", "B22, closes 18:15")
ctx.add_step("opened booking")
ctx.add_step("saw delay notification")
ctx.add_step("tapped 'Is my flight on time?'")
ctx.set_intent("task", "decide whether to leave for the airport now")

# traveller asks a question — agent already knows the full picture
messages = [ctx.as_message(), {"role": "user", "content": user_input}]

ctx.as_message() renders to:

Scope:
  user id: traveller-301

Facts:
  flight: BA 442 — London Heathrow to Amsterdam
  scheduled departure: 16:45
  status: Delayed — 2 hours (new departure: 18:45)
  gate: B22, closes 18:15

Steps:
  opened booking
  saw delay notification
  tapped 'Is my flight on time?'

Intent:
  task: decide whether to leave for the airport now

Demo

A traveller's flight is delayed. They open the app and type:

"Do I need to leave now?"

Your app has been watching. It knows which flight they're on, the new departure time, and when the gate closes. That gets passed to the assistant as background:

Facts:
  flight: BA 442 — London Heathrow to Amsterdam
  scheduled departure: 16:45
  status: Delayed — 2 hours (new departure: 18:45)
  gate: B22, closes 18:15

Steps:
  opened booking
  saw delay notification
  tapped 'Is my flight on time?'

Intent:
  task: decide whether to leave for the airport now

Without background the assistant can't answer at all.

python examples/evaluation.py
# → examples/evaluation_report.html

The model

Context has four sections:

Updating context

Use individual setters for explicit writes:

ctx.set_fact("current_page", "settings reference")
ctx.add_step("opened settings reference")
ctx.set_intent("task", "understand a configuration error")

Use merge helpers when multiple app surfaces contribute to the same section:

ctx.merge_facts({
    "docs": {"current_page": "settings reference"},
    "errors": {"current_error": "ImproperlyConfigured"},
})

Use merge() to combine two independently built contexts:

ctx.merge(other_ctx)

Exports

ctx.render()      # plain text for prompts
ctx.as_message()  # {"role": "system", "content": ...} for chat APIs
ctx.to_dict()     # plain Python dict
ctx.to_schema()   # validated Pydantic model

Validation

Optional. Subclass ContextSchema to add Pydantic checks:

from flowstate import ContextSchema

class DocsContextSchema(ContextSchema):
    facts: dict[str, str]

validated = ctx.validate(DocsContextSchema)