Activity Recorder
The Activity Recorder is what makes Ghast remember what you were doing, not just what you typed in chat. It is a background local recorder that samples your screen, extracts text on-device with macOS Vision OCR, throws the screenshots away, and keeps the redacted text and a few summaries in an indexed store the agent can search.
Two things define its design:
- Local. Capture, OCR, indexing, summaries all happen on this device. Nothing leaves unless you opt into 0G Storage sync for the activity category.
- Text-only by default. Screenshots are used as OCR input and discarded after analysis. Visual history (keeping the screenshot files) is off by default and only useful if you specifically want a visual archive.
The Settings page calls it Activity Recorder. The subtitle reads: Locally extracts screen text and builds activity memory. Screen images are discarded by default; data stays on this device.
What gets captured
When the recorder is on, it samples:
- Front app — the frontmost macOS application and its bundle id.
- Frontmost window title — usually the document, URL, or panel name.
- Browser tabs — visited URLs and titles from supported browsers.
- Screenshot → OCR text — a screenshot is taken at each sampling tick, OCR'd, and then discarded (unless you turned on visual history).
- Activity segments — contiguous "what you were doing" segments, opened and closed using an idle threshold. A segment groups events that belong to the same focused stretch of work.
What it does not capture:
- Keystrokes or typed content
- Clipboard contents
- Mouse movement
- Audio
- Continuous screen video
Activity segments
A segment is the recorder's unit of memory. It opens when you start doing something, accumulates events (focus changes, OCR text, browser tabs) while you keep working, and closes when:
- You idle past the idle threshold, or
- You lock the screen, or
- A muted (blacklisted) app comes to the front, or
- The recorder is disabled.
When the agent recalls "what was I working on at 3pm yesterday?", it's pulling a segment.
Status states
The status panel shows one of:
| State | Pill | Meaning |
|---|---|---|
| Off | disabled | Recorder switched off. Nothing being sampled. |
| Idle | idle | Switched on but not currently extracting. Ready to record. |
| Indexing | indexing | Recording — sampling, OCR-ing, indexing. |
| Screen locked | screen locked | Auto-paused while the screen is locked. |
Alongside the state pill the panel shows four metrics: Front app, Text index (size), Segments, Visual history.
OCR (text extraction)
OCR is the core of the recorder. Each capture goes through macOS Vision OCR locally; nothing is sent to a cloud OCR service.
Configurable knobs:
- Extract text from screen captures — master toggle for OCR. macOS only.
- OCR languages — Vision language codes, comma-separated (e.g.
en-US,zh-Hans). - OCR every Nth capture — skip OCR on intermediate frames to reduce database load (
1= every capture,5= one in five).
Text is redacted before being used for search and memory (sensitive patterns stripped at the recorder's text-handling layer).
Sampling controls
How aggressively the recorder samples is configurable under Sampling:
| Knob | What it does |
|---|---|
| Heartbeat interval (ms) | Force a capture this often even when nothing changes. |
| Idle threshold (ms) | Close an open segment after this much continuous idle time. |
| Minimum capture gap (ms) | Floor on how rapidly captures can happen — debounces busy moments. |
| Visual-change poll (ms) | Detect animations / video on this interval. 0 disables. |
| JPEG quality | 1-100; controls the quality of the temporary capture buffer (lower = smaller files). |
Most of these update live while the recorder is running; segment / idle timer changes apply on the next segment.
Privacy: storage policy and app blacklist
Two things keep the privacy surface small.
Storage policy
By default, screenshot files are not retained. The flow is: take screenshot → OCR → discard image → keep redacted text + summaries + indexes.
You can opt into keeping the screenshots:
- Keep visual history — off by default. Turn on only if you want Ghast to keep screenshot files after OCR.
- Max visual storage — cap on retained screenshot bytes. Oldest screenshot files are auto-deleted past this cap.
The text index, segment summaries, and embeddings are kept regardless.
App blacklist
Some apps you do not want recorded. The blacklist lets you list bundles (password managers, banking apps, anything sensitive) — when one of them is frontmost, the recorder auto-pauses and drops the in-flight capture buffer.
You can add apps two ways:
- Add from Applications… — picker that reads the app's
Info.plistfor the right bundle id. - Advanced: paste bundle id — when you want to type one directly.
The recorder treats the blacklist as defence-in-depth: bundles on the list never enter the database in the first place, even if the screen was technically captured.
macOS permissions
The Activity panel surfaces permission state with a Ready / Needs access indicator next to Screen Recording. Without screen recording permission the OCR pipeline has nothing to read.
Permissions are granted in System Settings → Privacy & Security → Screen & System Audio Recording. The panel deep-links to the right place.
Recent segments and danger zone
- Recent segments — a list view of the most recent segments with duration, event count, and a preview. Click into one to see its OCR text and summary.
- Danger zone → Delete everything — removes all segments, events, OCR text, summaries, embeddings, and any optional visual history files. The recorder configuration itself stays in place. Cannot be undone.
Three downstream uses
Once a segment is closed and summarised, it feeds:
- Daily summaries — "Today you mostly worked on the Ghast Desktop repo, read three papers on TEEs, and had a long chat with the team in Discord."
- Open-loop reminders — "You left a draft email open in Mail this morning — still want to send it?" Anchored to real activity, not generic nags.
- Memory-recall context — when you ask "what was that repo I looked at yesterday?", the recorder's index has the answer.
Activity-derived records are tagged with source: activity in the memory store so you can filter or exclude them in Settings → Memory.
What the agent sees
The agent does not get a raw stream of your activity. It sees:
- The daily summary when relevant.
- The matched record when a memory-recall query lands on an activity-tagged record.
- A specific segment's OCR text when you ask, "what was I looking at at 3pm?"
It cannot browse your activity freely without an explicit prompt that asks it to.
Disabling and clearing
| You want… | Do this |
|---|---|
| Stop new captures | Toggle the master switch off. Existing data is preserved. |
| Stop new captures from a sensitive app | Add it to the privacy blacklist. |
| Delete some segments | Use the per-segment delete in Recent segments. |
| Delete everything | Danger zone → Delete everything. |