All posts
Release2026-08-28

Chatbox 1.23: Memory, and a Work Mode That Keeps Up

1.23 fixes three things that kept getting in the way: the AI remembers you across chats, Work Mode is easier to steer, and long jobs are less likely to forget what they just did.

Memory: a new chat that still knows you

Every new chat used to start from zero. Chatbox now keeps the facts worth keeping, and brings them into the next conversation.

For exampleYou said weekly reports should be in Chinese, conclusion first, numbers second. Next week you open a fresh chat to write the report — it already knows the format.

Memory is split in two:

  • Global Memory — the default across chats, except for copilots with their own memory enabled. Writing taste, weekly habits, and the projects you keep coming back to.
  • Copilot Memory — enable it on a copilot, and its own memory replaces Global Memory for those chats. Your code-review copilot will not pick up your report template; your translator will not inherit repo conventions.

It works in Chat Mode and Work Mode. Toggle, review, and delete memories in Settings; the Work Mode panel has a quick switch too. On desktop you can scan local Claude / Codex memories and import only what you confirm.

Agent

Configure default agent behavior, its Soul, and its memories.

Memory
Save facts from conversations and recall them in new chats. Works in both Chat and Work Mode.
Memories3/100Scan again

Facts saved from conversations are loaded into new chats.

Weekly reports in Chinese. Structure: takeaway → numbers → risks.
Compare options in a table, not a long paragraph.
Deploy window is Tuesday evening, Beijing time.
Copilot Memories2
Code Review
PR titles in this repo use Conventional Commits.
Tests live in __tests__, not *.spec.ts.
Global Memory and Copilot Memory in Settings

Give Work Mode a Soul

In Work Mode the AI works from one global Soul — persona, tone, and boundaries. It takes the place of per-chat system prompts in Work Mode, so you are not rewriting “who you are” in every session.

For exampleYou write “Lead with the answer. Skip the fluff. Don’t say something is done if it isn’t.” From then on, whether it is cleaning a spreadsheet or ripping out dead code, it shows up with that personality.

Edit Soul under Settings → Agent. After you save, chats already running keep the snapshot they started with; new chats pick up the update. Changing the personality mid-task will not yank the current job off course.

Agent

Configure default agent behavior, its Soul, and its memories.

Soul

Soul defines who your agent is — persona, tone, and boundaries. It is loaded when an agent session starts.

# Soul

## Personality & Tone
Lead with the answer. Skip filler like “Great question!”

## Principles
Check existing files before you start. Don’t claim something is done if it isn’t.

## Boundaries
Ask before sending email or posting in public. Private things stay private.
SaveReset to template
The Soul editor

Work Mode, easier to live with

Work Mode is still the AI that does the work: run code, read and write files, call tools. 1.23 smooths the parts you touch every session.

Approval and folder, right on the input

New chats remember the working directory and code-execution permission you used last time. Both show up above the input, so you can see the current choice and change it without digging through the panel.

For exampleYesterday you edited reports in sales-dashboard with Smart Approval. Today’s new chat still shows those two chips, plus “Same as last time.” Switch folders or permissions by tapping the chips.
Smart Approvalsales-dashboardSame as last time
Type your question here...
Work Mode
Chatbox AI 5

AI-generated content may be inaccurate. Please verify important information.

Approval and working directory above the input

Queue a follow-up — or cut in

Send another message while Work Mode is still going, and it lines up behind the current reply. You can edit or drop items in the queue. Need to redirect right now? Hit Send now and it interrupts the current turn to hear you first.

For exampleThe AI is totaling revenue for the whole year. You realize “leave Q3 alone — Q4 only.” Drop it in the queue, or hit Send now and turn the ship immediately. No waiting for a finished answer you already know is wrong.
Will send after the current response finishes1 / 20Clear queue
1Leave Q3 alone — Q4 onlySend now
Type your question here...
Work Mode
Chatbox AI 5
Queue a message during generation, or cut in

Full Access no longer pauses every 25 steps by default

Code execution has three levels: Always Ask, Smart Approval, and Full Access. On Full Access, commands and file changes skip per-action approval — and by default, long runs no longer pause every 25 steps. Use it when you already trust the job to finish on its own. If you turn “Pause after every 25 steps” back on in the conversation settings, that chat will still pause as requested.

For exampleOvernight data cleanup. Switch to Full Access and leave “Pause after every 25 steps” off, so you do not have to wake up to click Continue. Everyday tasks can stay on Smart Approval.

With a vision-capable model, it can go look at pictures itself

With a vision-capable model, Work Mode can open images on its own — a file in the working directory, or something it just downloaded — then revise copy, check a chart, or write from what it sees.

For example“Compare assets/hero.png with a reference image from the web and shorten our headline.” It opens the local image and downloaded reference instead of waiting for you to paste screenshots.
View Image· /project/assets/hero.png
Run Command· curl -L https://picsum.photos/1200/630 -o /project/reference-hero.jpg
View Image· /project/reference-hero.jpg
Chatbox AI 5 · 428 tokens
Work Mode opening a local image and one it just downloaded

Long jobs that remember the last step

The longer an agent runs, the more tempting it is to trim context too hard. 1.23 changes two cleanups so the model can still pick up what it just did.

Tool results are not cleared too early

Older tool results used to disappear too early. The model remembered that it had read a file, but not what was in it. Now Chatbox waits until context is genuinely tight, then replaces older payloads with a compact cleared marker. The tool call remains, so the model can run it again if it needs the details, while the newest result stays intact.

For exampleIt just read schema.sql and is changing an API. Field names are still there, so it does not have to re-read the file — and is less likely to invent a column.

Summaries that match the conversation, with a recent raw tail

When a chat gets long enough to compact, Chatbox first figures out whether you are working a task, writing a story, or just talking, then summarizes with the matching outline. When the context budget allows, up to the last two conversation rounds stay verbatim. If the recent content is unusually large, Chatbox shortens that raw tail so compaction can bring the chat back within the model window.

For exampleIn an ordinary text-only thread, forty turns go into a login-page rewrite. After compaction, the earlier exploration becomes a briefing, while up to the latest two rounds remain intact: “make the primary button green, and bump the type size.” It keeps going without making you restate the request.
Earlier exploration summarized: goal, files touched, what remains
Earlier messages summarizedView
Make the login button green, and bump the type size.
On it. I’ll update the primary button in the header to brand green and take the type from 14 to 16.
Compaction keeps up to two recent rounds intact when the context budget allows

More models, ready when you are

1.23 adds Tencent Hunyuan, Xiaomi MiMo, LongCat, GLM Coding Plan, and OpenCode Zen / OpenCode Go. Web search can also use a self-hosted SearXNG instance.

Tencent HunyuanXiaomi MiMoLongCatGLM Coding PlanOpenCode ZenOpenCode GoSearXNG

Try it when stable 1.23 arrives

Once stable Chatbox 1.23 is released, update to use these features. Memory will work across platforms; Work Mode, Soul, and View Image with a vision-capable model will be available on desktop (Windows / macOS / Linux).

Read the Work Mode guide

Also in 1.23: delete all archived chats, preview text files from tool results, and support the 2026-07-28 MCP protocol for custom servers.