Set up SPARK and follow its learning

This walkthrough starts with a small, supervised SPARK trial. You will authorize activity collection, inspect what was captured, let a selected model review it, and understand the difference between a candidate, a memory and a proactive message. You do not need to enable everything at once.

Use a separate test environment if you intend to create artificial personal facts. A successful setup means the controls and processing work; it does not mean SPARK should immediately learn something about you or send a suggestion.

1. Check the three prerequisites

Before allowing collection, open the following screens:

1. Models: confirm that the model you want to use is configured and can answer an ordinary request. SPARK’s learning model need not be the same model you use in your main conversation.
2. Memory → Setup: configure a working embedding model for long-term memory. Learning depends on this service. If you skipped it during onboarding, follow the memory guide first.
3. Home → Computer access: follow the connection or local access instructions for your deployment. When a companion connection is required, wait for the connected/ready status (Connected on Mac) rather than treating a completed download as a connection.

The browser’s operating system is not necessarily the Runtime’s. A Docker Runtime needs access to the native desktop through its computer connection; a WSL process does not itself read Windows application content. Follow the platform instructions, including any OS accessibility permission or companion-update notice.

In SPARK’s Settings, expand Computer connection to inspect Computer and Collection access. Manage Computer access takes you back to setup. A connected computer can still lack accessibility permission or readable content in a particular application.

2. Authorize collection without starting it

Open SPARK → Settings → Activity permissions. There are three cards:

Card	Permission	Purpose
Collection	Allow collection	Notice eligible changes in active applications.
Learning	Allow learning	Review retained activity and propose useful knowledge.
Proactive mode	Allow proactive mode	Consider sending a useful message in SPARK chat.

For the first trial, allow collection and learning but leave proactive mode off. Choose a Learning model before saving. Nothing is gained by enabling unsolicited messages before you have checked what the collector and learning model produce.

When collection permission is first granted or collection is started, ABot shows Before allowing activity collection. Read the explanation and the linked Software Terms and Privacy Policy. The acknowledgment checkbox starts unchecked; Continue is available only after you select it. Cancel leaves that acknowledgment unapproved.

The acknowledgment is saved for the installation, shared across its environments and kept across updates. It is not itself a Start command. Save settings stores the permitted activities and settings; the separate Start/Stop controls determine whether they run. Previously enabled collection can resume after a Runtime restart according to saved settings, so explicitly stop it when ending an experiment.

3. Choose when learning runs

In the Learning card, Start a review offers two choices:

• By time: use Review interval (minutes) when you want pending activity considered at intervals. An interval does not force a model call when there is no eligible observation work.
• By observation count: use Eligible observations waiting when you prefer to accumulate activity before reviewing it. The number counts eligible observations, not minutes, words or distinct applications.

For a short supervised trial, you could choose By time with an interval of 15 minutes. This is a suggested starting value, not a requirement or a promise that a review finishes at that exact time. Longer intervals reduce review frequency; shorter ones can consume the shared call budget sooner. With observation-count mode, storage pressure can start eligible work earlier than the chosen count. Hours, pauses and limits still apply.

Each activity has its own Set activity hours checkbox. Enabling it reveals From, Until and Time zone. Enter an explicit zone such as Europe/London or your own IANA zone, and make sure the current time lies within the window during your test. Collection hours do not automatically become learning or proactive hours.

4. Set limits before leaving it running

Under Shared limits, inspect these fields:

Field	How to use it
Model calls per day	Shared allowance for learning and proactive reviews; failed attempts count too.
Proactive messages per day	Maximum proactive deliveries, not the number of review calls.
Concurrent cloud calls	Limits eligible cloud work; local models still run one call at a time.
Embedding calls per day	Separate allowance for embedding operations.
Embedding input characters per day	Bounds daily text sent for embedding.
Daily budget time zone	Determines the budget’s daily reset boundary.

The form’s starting values include 24 model calls and 96 embedding calls per day. Review the actual saved values in your installation. These are not currency or token spending caps. One learning review can require multiple model calls, especially with the Supervisor route, so 24 calls does not necessarily mean 24 reviews.

Keep concurrency at 1 for an initial trial. After observing actual use, adjust limits deliberately rather than increasing everything when processing pauses. Changing the budget time zone does not reset the current allowance. Select Save settings and check the feedback before leaving the form.

5. Capture a small amount of non-sensitive activity

Prepare a plain document containing harmless test text in an accessible editor, or open a public page. Close sensitive applications first. The Applications list is populated as activity is collected; it is not an inventory that lets you preselect every installed application before the first capture.

1. Select Start collection. If the button says Review collection access or Set up collection, complete the missing step instead.
2. Switch to the non-sensitive application and make a small text change or move between a few public pages.
3. Return to SPARK and open Activity.
4. Check Last collected and Pending observations. Use Refresh if you need to request the current state.
5. Select Stop collection once you have enough activity to inspect.

Collection uses text exposed by accessibility APIs after relevant changes settle. A quiet, unchanged screen does not generate periodic screenshots. Repeated content can be deduplicated; some revisits have no new text. Protected controls and inaccessible canvas content may not be readable. The collector does not record video, audio, clipboard history or keystroke history, and does not perform OCR.

6. Control applications and review the evidence

Open Applications. Each observed application has two independent switches:

• Collect activity controls future captures from that application.
• Process activity controls whether its retained observations can be reviewed.

Turning off processing leaves retained activity pending within storage and expiry limits. Neither switch retracts data already sent to a provider or deletes saved memories. If something sensitive was captured, stop collection and use Activity → Delete pending… to remove queued activity after confirmation; review established memories separately.

Select Start learning when the model, permissions and limits are ready. In Activity → Recent reviews, select a batch to open Review details. Expand an observation to read Captured text, its application/window identity and Capture details where available. The summary distinguishes saved records from candidates.

Waiting means a batch has not completed review. Reviewing is active work. Completed or Candidates updated does not guarantee a new lasting memory. Failed provides a reason; expand Technical details when needed. Inspect the evidence before treating an inference as true: text being visible does not prove you authored, read or agreed with it.

7. Inspect candidates before expecting memory

Open Knowledge. The Candidates column contains provisional records from conversations and computer activity. Why this candidate shows its explanation, independent reinforcement count and dates. A candidate is not used as established memory yet, even if its first score is high.

Expand Settings → When a candidate becomes memory to inspect Promotion score (out of 100), Days without reinforcement, Maximum candidates and Candidate storage limit (MiB). Start with the existing policy rather than lowering the threshold to force your experiment to succeed. Promotion also requires independent evidence; repeated copies of the same screen do not establish a lasting fact. Scores are model judgments, not probabilities.

The Established memory column shows recent records. Select Manage memories for the full list, editing and deletion. Conversation history, raw pending activity, candidates and stored memories are separate things. Clearing one does not clear all the others.

8. Enable proactive chat only when ready

Return to Settings → Proactive mode. Choose Chat and proactive model, or Same as learning if that is what you intend. Set Check interval (minutes), optional activity hours and Proactive messages per day, then save. Select Start proactive mode separately.

A check may decide there is nothing useful to say. Proactive mode can use established knowledge while collection is off; turning off collection is therefore not the way to stop messages.

In the Chat tab, choose Permissions and select Save. Begin with Ask while testing. The Thinking mode selector depends on the selected model’s reasoning support. Open chat opens the persistent SPARK conversation; its settings button returns to these controls. Chat uses the model selected in the Proactive mode settings.

Background learning and proactive reviews cannot invoke tools or execute their suggested actions. A message you submit in SPARK chat is different: it is an ordinary request using that chat’s permissions. A saved suggestion may offer Open a session about this. That action opens a regular conversation with retained context but submits no request; review it before asking ABot to act. Older suggestions without saved context explain why their handoff is unavailable.

9. Diagnose a stalled trial and stop cleanly

Open Resources → Today’s resource use to see model calls, embedding use, pending observations and the next reset. Displaying these counters does not itself run the model.

Symptom	Check first
No observations	Collection running state, activity hours, computer readiness, OS access and whether the app exposes text.
Pending count does not decrease	Learning’s Start state, eligible count/interval, processing exclusions, memory setup and shared limits.
Learning pauses during chat	Interactive requests take priority; allow background work to resume afterward.
No candidate or message appears	A completed review may find nothing durable or useful. Inspect the review, not only its count.
Memory seems wrong	Read its source evidence, then edit or delete it in Memory. Do not solve this by raising usage limits.

To end the trial, use Stop collection, Stop learning and Stop proactive mode as applicable. Stops preserve authorization and settings. Revoking the corresponding Allow permission also prevents restarting until permission is granted again.

Raw pending evidence has a 24-hour expiry and bounded storage; cleanup runs while the service is active and at startup, not while the process is stopped. Delete pending… is a separate immediate confirmed action. Chat → Clear messages… clears that transcript while keeping memory and permissions. Review or delete established records in Memory separately.

Finally, remember the processing boundary: hosted learning, proactive or embedding providers can receive relevant content. Recognizable secrets are redacted from newly collected fields, but unknown formats can escape detection. Local storage and redaction are not a guarantee that sensitive information stays on your computer.