report

immich-memories report prints a report of one run that you can paste into a GitHub issue. It runs on your machine, makes no network request and sends nothing: you read it, then decide what to share.

immich-memories report                 # the latest run, as Markdown
immich-memories report <run-id>        # one run
immich-memories report --json          # the same, for tooling
immich-memories report --bundle report.zip   # the full report and its logs, as a ZIP

The report holds the run's settings as shape (keys, not values), its errors and warnings, the log, and a table of where the time and memory went, phase by phase (peak MB for the process, and with ffmpeg added). That table is the first thing to read when a cut is slow: immich-memories runs show prints the same timings in full, per picture.

Before anything reaches the report, it removes credentials, the names of the people you know, albums and places, GPS coordinates, IP addresses, hostnames, URLs and absolute paths. IDs become hashes that only match inside that one report. Pictures are never included. Exactly what is removed, and how: Privacy.

A run of generate --ask (a film from a sentence) adds a free-text section: the sentence, the translation trace, the translation as data (the same JSON --ask-trace writes: each part, the pool's photo and video counts, the verdict), the pool funnel and the engine's picks as hashed IDs. The trace gets its own block in the Markdown, so a long one is pasted whole. A film rendered later from a saved cut (runs render, or Render in the web UI) is a run of its own, and its report carries the request its cut was made for, with the render's picks. Names become roles ("the owner's son"), place names become "area A", words read by OCR become "text-1", and a birth date the trace dated from becomes [private]. Captions stay out unless you pass --include-flagged-captions.

When the film got it wrong, say so on the run, then paste the report:

immich-memories report <run-id> --wrong <asset-id> --wrong <asset-id> --missing "the farm gate"

Each photo marked wrong gets the step that let it into the pool and whether the engine picked it. The missing words are checked against the run: did the reading keep them, were they offered or picked as the subject, and does any caption in the pool say them. The marks stay on the run, so a later report shows them too.

The web UI has the same report behind Copy report on a run's page, and the --bundle ZIP behind Download report. Every flag is in the CLI reference.