See what your agent just remembered: /memoir:pane

The problem: memory you can't see

Memoir's Claude Code plugin captures memories on its own. When a turn ends, the Stop hook reads the transcript, decides what is worth keeping, and commits it to the store. It runs in the background so it never slows the session down, which also means its writes land a few seconds after the turn, with nothing on screen to say they happened.

Finding out what was stored meant going to look: /memoir:status, the CLI, or the web UI. Most people didn't, and memory nobody looks at is memory nobody corrects. A wrong preference captured on Monday is still steering the agent on Friday.

Branches add a second question. Memoir gives every code branch its own memory branch, so a feature's captures stay isolated until you promote them. Before promoting, you want to know what is different from main, and there was no quick way to see that either.

A count under the prompt

The smallest part of the change is one line. The hint line under the prompt now ends with the store's state:

memoir: feature/auth · 18 memories · +2 this turn

+2 this turn appears once the turn's captures have landed and clears when the next turn starts. Metric writes don't count toward it. There is nothing to configure: the old status-line widget needed a hand edit to ~/.claude/settings.json, which most people never made. If you did make it, the plugin notices and skips its own line so the state isn't shown twice.

The pane: three tabs

/memoir:pane opens a pane. In a wide fullscreen terminal it docks beside the transcript; otherwise it sits above the prompt. ctrl+x tab focuses it and ctrl+x x closes it.

• Memories (the default) answers the branch question. On a feature branch it lists only the keys that differ from main, each marked + (new on this branch), ~ (changed here) or - (on main only, forgotten here), with one line of its value. Open a changed key and main's value sits beside the branch's. From there you can forget the key on this branch or restore main's value, each behind a confirmation. p merges the whole branch into main and asks how to settle keys changed on both sides: keep main's value, take the branch's, or keep both by appending one to the other. A filter: field narrows the list, and t shows the taxonomy above it.
• Recent is the capture feed: the store's last 20 or so commits, newest first, with ● marking the ones from this turn. Open one to see what it changed, each key's value before and after. Metric writes are hidden until you press m.
• Metrics draws one bar chart per metric across your branches (turns, tool calls, tool errors, average latency, output size), with the current branch first. The figures are the same ones the web UI's Metrics tab shows.

The pane only writes what you confirm in it, and it writes through the same memoir CLI the hooks use. Nothing about how capture works has changed; the four shell hooks, the skills and every existing command are untouched.

Getting it

• Update the Memoir plugin and run Claude Code 2.1.287 or newer. The count under the prompt appears on its own; /memoir:pane opens the pane.
• On older Claude Code versions, /memoir:pane prints a plain-text summary of the store and the latest captures instead.

See the pull request on GitHub

Auto-capture earns trust only when you can watch it work. Now it happens on screen, one turn at a time.