Step 5 — Enrich with user memory

Rich user memory — mastered, in_progress, best_completion_pct, journey_completion, and related fields — is expressed in terms of named workflows. Those names only make sense once you have supplied a workflow ontology (which journeys exist and what “done” means). Treat ontology definition as the input step: Autoplay Atlas labels replays by workflow and completion; distillation builds the memory profile your onboarding agent reads.

Support for this step is coming soon.

Define your workflow ontology

Supply the ontology of workflows your product recognises — which journeys exist, how they relate, and what steps matter. Session replays can only be labelled against workflows that exist in that ontology. Golden paths and recorded flows are how that ontology gets into Autoplay; they are not an end in themselves. You can establish workflows in two ways:

Chrome extension — record flows directly in your UI with the Autoplay Chrome extension.
Documentation → workflows — convert existing product documentation into structured workflows with Autoplay. Join our Discord and reach out to the team to get started.

Use one or both; coverage of the journeys you care about matters more than which channel you use first. Iteration with Autoplay on AI accuracy: production-grade labelling is a collaborative loop, not a one-shot upload. You will work with the team to iterate on AI accuracy for workflow assignment and completion signals on session replays — especially when ingesting workflows from documentation.

Autoplay Atlas and session replay labelling

Once your ontology exists, Autoplay Atlas labels session replays by workflow and workflow completion rate — replay-level semantics grounded in those workflows. For each replay you typically see:

Workflow completion rate — how far the user progressed on the workflows you defined, within that replay.
Most likely workflow — which journey best explains what they were trying to do in the session.
Timeline — how inferred workflows and completion evolve across the replay, not only a single summary at the end.

That turns a flat stream of clicks into workflow-level context you can feed to an agent. Individual UI events are ambiguous; Atlas lifts them into which workflow, how complete, and where they diverged — so the model reasons about journeys, not isolated DOM events.

Why workflow context beats raw clicks alone

Raw telemetry is a list of atomic actions: without grouping, you cannot tell exploration from a stuck multi-step task. When Autoplay matches the user’s sequence against your defined workflows, you get structured outcomes — for example which workflow was attempted, how complete the attempt was, and which expected step was missed. Completion rate then tells the onboarding agent how to intervene (introduce a flow, resume mid-way, or stay quiet after mastery). Cross-session memory builds on the same workflow signals so progress is not forgotten between visits.

Index journeys for retrieval

Index ideal step-by-step journeys in Autoplay — from docs and/or captures via the Chrome extension — so they can power retrieval and Atlas-backed labelling as those services roll out.

Calling query_knowledge_base from your app matches the Knowledge base SDK — that helper and HTTP endpoint are coming soon. Supplying workflows via the extension or with the team is separate from wiring this query path in your codebase.

golden_paths = await client.query_knowledge_base(
    query=payload.to_text(),
    top_k=2,
)

See Knowledge base for the planned endpoint reference and golden path management.

Ontology + live actions

With real-time events plus a workflow ontology (and Atlas labels on replays), your agent can reason over which workflow someone is in and how complete it is — then compare that to live actions. Once the knowledge base query ships in your stack, you can retrieve ideal-path text at inference time:

async def on_actions(payload: ActionsPayload) -> None:
    golden_paths = await client.query_knowledge_base(
        query=payload.to_text(),
        top_k=2,
    )

    suggestion = await your_llm(
        system="You are a product onboarding agent. Use the ideal path and live actions "
               "to suggest one concrete next step. Be brief.",
        user=f"## What the user is doing\n{payload.to_text()}\n\n"
             f"## Ideal paths for this context\n{golden_paths.to_text()}",
    )

User memory profile

Coming soon.

Once ontology and labelling feed distillation, fetch the user’s cross-session memory profile alongside the live event payload. This profile is built automatically by Autoplay — after every session, an LLM distillation pass updates it with what the user did, what they completed, and where they struggled.

memory = await client.get_user_memory(user_id=payload.user_id)

The profile tells you exactly where this user is in the product’s adoption journey — including workflow completion fields tied to the ontology above:

{
  "knowledge_state": {
    "mastered": ["connect_data_source", "create_dashboard"],
    "in_progress": {
      "setup_slack_integration": {
        "attempts": 2,
        "last_issue": "missed final step: verify connection success message",
        "best_completion_pct": 85.0
      }
    },
    "untouched": ["set_up_alerts", "configure_webhooks", "export_to_csv"]
  },
  "journey_completion": {
    "completed_workflows": ["connect_data_source", "create_dashboard"],
    "total_known_workflows": 7,
    "completion_rate": 0.29
  }
}

What your onboarding agent can do now

User memory is the relevance filter. Without it, your onboarding agent treats every user as if it’s their first day. With it, suggestions are grounded in exactly what this person has and hasn’t done:

Memory state	Onboarding agent behaviour
Workflow in `mastered`	Skip it — the user already knows it
Workflow in `in_progress`	Surface the specific missed step, not the whole flow again
Workflow in `untouched`	Proactively introduce it when the user is in a relevant context
Area in `struggle_patterns`	Change approach — repeating what already failed is noise

Code — memory alone

async def on_actions(payload: ActionsPayload) -> None:
    memory = await client.get_user_memory(user_id=payload.user_id)

    # Don't suggest what they already know
    mastered = memory.knowledge_state.mastered
    gaps = memory.knowledge_state.untouched + list(memory.knowledge_state.in_progress)

    suggestion = await your_llm(
        system="You are a product onboarding agent. Only suggest things the user hasn't mastered yet.",
        user=f"## What the user is doing\n{payload.to_text()}\n\n"
             f"## Already mastered (do not suggest)\n{', '.join(mastered)}\n\n"
             f"## Gaps to focus on\n{', '.join(gaps)}",
    )

See User memory for the full schema and field reference.

Combine all signals

With real-time events, user memory, and golden paths together, your onboarding agent can compute a precise gap: where they are now, what they have already done, and where they should be going next.

Where the user is now        ← real-time events
What they've already done    ← user memory (mastered / in_progress)
Where they should be going   ← golden path from knowledge base
                                       ↓
              gap = ideal journey minus what this user has completed
                                       ↓
         onboarding agent surfaces the single most relevant next step

import asyncio

async def on_actions(payload: ActionsPayload) -> None:
    memory, golden_paths = await asyncio.gather(
        client.get_user_memory(user_id=payload.user_id),
        client.query_knowledge_base(query=payload.to_text(), top_k=2),
    )

    mastered = memory.knowledge_state.mastered
    in_progress_keys = list(memory.knowledge_state.in_progress)
    context = f"""
## 👆 What the user just did
{payload.to_text()}

## 📜 Their workflow history
Mastered: {', '.join(mastered) or 'none'}
In progress: {', '.join(in_progress_keys) or 'none'}
Never started: {', '.join(memory.knowledge_state.untouched) or 'none'}

## 🗺️ Ideal path for this area
{golden_paths.to_text()}
"""

    suggestion = await your_llm(
        system="You are a proactive product onboarding agent. Based on what the user is doing, "
               "their history, and the ideal path, suggest one concrete next step. "
               "Do not suggest anything they have already mastered. "
               "Be brief and specific. Stay silent if there is no clear gap.",
        user=context,
    )

Recipe overview: The proactive onboarding and adoption agent gold standard

Who We Are

Get Started

Live User Activity

AI Support Agents

Visual User Tours

Receive events

🔌 Live activity via MCP

Build with events

Integration helpers

Context sources

Reference

Step 5 — Enrich with user memory

Define your workflow ontology

Autoplay Atlas and session replay labelling

Why workflow context beats raw clicks alone

Index journeys for retrieval

Ontology + live actions

User memory profile

What your onboarding agent can do now

Code — memory alone

Combine all signals

​Define your workflow ontology

​Autoplay Atlas and session replay labelling

​Why workflow context beats raw clicks alone

​Index journeys for retrieval

​Ontology + live actions

​User memory profile

​What your onboarding agent can do now

​Code — memory alone

​Combine all signals

Define your workflow ontology

Autoplay Atlas and session replay labelling

Why workflow context beats raw clicks alone

Index journeys for retrieval

Ontology + live actions

User memory profile

What your onboarding agent can do now

Code — memory alone

Combine all signals