Docs

Appearance

CharacterVault 1.8.5 ​

CharacterVault 1.8.5 warns before you leave a running Agent and marks runs that leaving cut short, adds an option for the Agent to remember exactly what it wrote earlier in the chat, and has it pick real names that fit the setting, plus em dash matching that leaves hyphens alone and a large round of Agent and chat fixes.

Try it: https://vault.charactervault.appStaging: https://staging.charactervault.appDocs: https://vault.charactervault.app/docs/

Highlights ​

Leave warning for Agent runs ​

A run's edits, and a review waiting for you, are only written when the run finishes or you apply the review. Refreshing or closing the tab before that now asks first, so you do not lose them by accident. There is nothing to turn on. With the Agent idle, the page leaves without asking.

  • Browser confirm: also when the tab is in the background or the window is minimized.
  • If you leave anyway, the edits are lost: that covers a refresh mid-run, a refresh with a review waiting, or going back to the library with a review waiting. When you reopen the chat, that run says Interrupted by leaving or refreshing the page. These edits were not saved. Its write lines show struck through as Not saved.
  • The Agent knows: ask "what did you change?" and it knows nothing from that run landed. Finished, stopped, and applied runs show no notice.

AI Agent → Leaving mid-run

Remember earlier edits ​

Settings → Character Workspace → Agent edits → Remember earlier edits in the chat is off by default. Turn it on and the Agent sees each earlier run's edits as the calls it made, so "reword the second sentence you just wrote" works without searching for it.

  • Off: the Agent gets a short note on what each earlier run changed, as before.
  • Both Agents: the character and lorebook Agent share the setting, and it covers runs from before you turned it on.
  • Limits: interrupted runs, runs saved before 1.8.5, and runs larger than a quarter of the input budget fall back to the short note. The context meter counts the replayed edits.

AI Agent → Remember earlier edits

Real names for new characters ​

Ask the Agent to rename the character or add NPCs without giving names ("add her two brothers, they grew up in 1920s Kraków") and it picks real names that fit the setting instead of Elara- or Voss-style names.

  • 44 origins: from Polish and Italian to Yoruba and Vietnamese. Japanese settings get romanized names (Sato, Yuko).
  • No repeats: never a name already on the card or in its lorebook.
  • Invented cultures: for "a desert clan" or "a forest elf village" it makes up names that share one sound across the group.
  • Both Agents: the character Agent and the standalone lorebook Agent.

AI Agent → Character names

Also in this release ​

  • Errors on your message: when an Agent request fails, the error shows under your message instead of in a hover tip.
  • Em dashes: an em or en dash in an Agent search or replace matches only itself. "Remove em dashes" on a card with hyphenated words (demi-human, t-shirt) leaves the hyphens alone, and says so when none are left.

Fixes ​

  • Agent chat: speech and thinking from lookup-only turns stay in the chat (thinking only when Show reasoning is on), a stopped or failed turn keeps the text that already streamed, and old errors and turn-limit notes no longer come back after a reload or a retry.
  • Older messages: deleting from the oldest visible message brings older saved ones back in the Agent and Orion chats, and Edit then send includes them.
  • Runs and Stop: a tool error or an unreadable tool call no longer loses the run's other edits, Send waits after Stop until the stopped run has saved, and Send, Retry, or Delete with a review waiting opens the review instead of removing messages under it.
  • Linked lorebooks: an Agent edit on a card linked to a library book no longer replaces that book and its other linked cards when the card was out of sync.
  • Small context windows: trimming a long chat keeps your current request, shortens the newest tool result first, and can drop a huge earlier request.
  • Review: --- scene breaks inside a greeting survive editing the Greetings list (the separator becomes ---- when needed), a deleted and an added entry stay two changes, and renumbered new entries link to their saved id.
  • Greetings: filling an empty greeting slot works in place, and the Agent will not empty a greeting.
  • Providers: think tags in non-streamed replies go to Thinking, tool calls from gateways that send no index stay apart, a 400 retry never drops your messages or model and only remembers stripped settings once the retry works, replies that mention <<< or <tool_call> are no longer cut, and stopping a long XML reply keeps only the speech.

Upgrade ​

Client-only app. Open the hosted site, rebuild locally, or rebuild the Docker image; no server migration.

bash
git clone https://github.com/spaceman2408/CharacterVault
cd CharacterVault
npm install
npm run dev

Running it with Docker? Pull and rebuild the image:

bash
git pull
docker compose up -d --build

Local data from 1.8.4 keeps working. Agent chats now save each write's arguments with the message, and the new setting is included in settings backups. Hard refresh if the UI looks stale after deploy.

Docs ​

Privacy

Copyright © 2026 spaceman2408