Trails organizes agent-heavy work around the questions you ask when memory gets foggy: What happened today? Where did the week go? What is still in motion?
Days
The day view is the primary reading surface. Each project gets one timeline lane:
- A pale wash shows when agents were active.
- Solid color marks show minutes when you were present and steering.
- The story below the timeline summarizes each project’s contribution to the day.
Use the previous and next controls to page through days. Trails uses a human-shaped day boundary: work after midnight can remain attached to the evening it grew from rather than being split at 00:00.
Week
The week view stacks the same timeline grammar across several days. It is designed for reconstruction, not evaluation: you can see which projects held your attention and how agent activity continued around it without turning the week into a score.
Attention time can roll up into day-credits for people who bill in quarters, halves, or full days. The readable week remains the source of truth; billing is a derivative of that memory.
Threads
Threads collect work that spans sessions and days. A thread can be:
- in motion — active work is continuing.
- waiting on you — the next useful move needs your input.
- resting — deliberately paused; not a failure state.
- dormant — quiet long enough to move out of the active set.
The language is intentionally judgment-free. Trails does not add deadlines, priority scores, streaks, or productivity grades.
Projects and engagements
Projects are discovered from session working directories and repository context. You can rename projects and assign related projects to a shared engagement. Those edits live in canonical SQLite on the hub Mac—your same working Mac in a one-Mac installation—so every browser sees the same organization.
When a summary is missing
Summary generation waits for changed sessions to settle before processing them. If the optional inference relay is unavailable, trails continues to work and uses the first prompt as a fallback. Timeline activity and normalized session observations do not depend on generated prose.