Data folder

All data lives in one folder, ~/.config/daily-recall by default, created readable by your user only. Settings › Storage can point the app at another folder.

~/.config/daily-recall/
├── settings.json     settings, hand-editable
├── recall.db         SQLite database (WAL mode; safe to read while the app runs)
├── recall.lock       held by the running app
├── shots/            screenshots, shots/YYYY-MM-DD/HHmmss.jpg
├── marked/           marked moment images, marked/YYYY-MM-DD/HHmmss.jpg
├── dictation/        scratch audio while dictating; deleted after each dictation
└── summary-runs/     the exact prompt and raw reply of every summary and digest request

Only one running Daily Recall uses a data folder. A second copy started on the same folder records nothing and offers to quit, to quit the other one, or to choose another folder.

What is kept

DataKept
Window spans, meetings, notes, clipsforever
Screenshot and marked-moment descriptionsforever
Summaries and digests, all versionsforever
Screenshot images14 days by default
Marked moment imagesforever by default
Summary request logsforever by default

Retention runs at launch, once a day and when you change a period. It deletes only files. A screenshot image is deleted only after its description exists.

Database migrations can’t be reversed. Once a newer version has opened the database, an older version can’t open it anymore. Keep a copy of recall.db before you try an older build.

Tables

TableOne row per
focus_spansstable foreground state: app, window title, URL, start, end, end reason
screenshots, analysescapture, and the model’s description with the exact prompt sent
meetings, meeting_digestsTeams call with your notes, and each digest version
notesnote, with title and Markdown body
web_clips, page_digestsclipped page per day, and each digest version
summariessummary version of a day
search_docs, search_ftssearchable item, with an FTS5 full-text index

Times are UTC milliseconds since the epoch. Each row also stores the local UTC offset at the time.

Query from the shell

Today’s windows in local time:

sqlite3 -header -column ~/.config/daily-recall/recall.db "
  SELECT time(started_at/1000, 'unixepoch', 'localtime') AS start,
         time(coalesce(ended_at, last_seen_at)/1000, 'unixepoch', 'localtime') AS end,
         app_name, window_title, url, end_reason
  FROM focus_spans
  WHERE started_at >= strftime('%s', 'now', 'localtime', 'start of day', 'utc') * 1000
  ORDER BY started_at;"

Today’s screenshot descriptions:

sqlite3 -header -column ~/.config/daily-recall/recall.db "
  SELECT time(s.taken_at/1000, 'unixepoch', 'localtime') AS time, s.app_name, s.status, a.response
  FROM screenshots s LEFT JOIN analyses a ON a.screenshot_id = s.id
  WHERE s.taken_at >= strftime('%s', 'now', 'localtime', 'start of day', 'utc') * 1000
  ORDER BY s.taken_at;"

Full-text search for a word:

sqlite3 -header -column ~/.config/daily-recall/recall.db "
  SELECT d.kind, datetime(d.started_at/1000, 'unixepoch', 'localtime') AS time, substr(d.body, 1, 80) AS text
  FROM search_fts JOIN search_docs d ON d.doc_id = search_fts.rowid
  WHERE search_fts MATCH '\"atlas\"' ORDER BY d.started_at DESC LIMIT 20;"

Why a summary failed

Every summary and digest request leaves two files in summary-runs/<day>/: the exact prompt sent, and the result with HTTP status, duration and the server’s raw reply. For example, "finish_reason": "length" with empty content means thinking used up the token limit. Turn off “Let the model think” for that prompt, or raise the server’s context.