Status beat leaks between bots

Support Bug Reports
, ,
1

Where does the bug appear (feature/product)?

Grok Bot

Describe the Bug

Two separate Grok Bot chats/bots are using the same underlying agent for different long-running research projects. Project A is running in one bot/chat, while Project B is running in a separate bot/chat.

Both projects require regular status beats every few minutes, reporting what was found and including a link.

The problem: progress from one project shows up in the other bot/chat. Because both bots share the same underlying agent, durable memory and conversation history contain information from both projects. A status beat for Project B keeps Project B “hot,” so the next turn in the separate bot/chat often answers with Project B findings instead of the findings for that bot’s own project.

The two bots are separate conversation surfaces, but they are backed by the same agent ID. The two projects can also have separate disk folders; the mix-up still occurs within the agent’s shared memory/history.

Steps to Reproduce

  1. Create two separate Grok Bot chats/bots that use the same underlying agent.
  2. Run Project A in Bot/Chat A.
  3. Run Project B in Bot/Chat B.
  4. Give both projects recurring status-beat routines that run every few minutes and report findings with links.
  5. Allow several Project B status beats to run, refreshing Project B in the shared agent memory/history.
  6. In Bot/Chat A, ask for Project A status, links, or an ambiguous request such as “continue.”
  7. Observe: Bot/Chat A receives Project B content instead of Project A content.

Expected Behavior

Each separate bot/chat should remain isolated to its own project, even when both are backed by the same underlying agent.

A status beat generated for Project B should not contaminate, refresh, or prime the context for Project A.

Durable memory and conversation history should be scoped to the individual bot/chat or project, rather than globally shared across separate bots using the same agent.

Actual Behavior

Bot/Chat A receives Project B digests.

The underlying agent’s durable memory/history contains both projects as standing facts. Recurring beats for Project B make the leak worse by continually refreshing Project B’s context, making it more likely to appear in the next turn of Bot/Chat A.

The two bots are separate, but they are sharing the same agent and therefore sharing the agent’s memory/history.

Operating System

Windows 11

Version Information

Not provided

For AI issues: add Request ID with privacy disabled

Shared underlying agent: 92b2d70e-b956-44b5-945b-849762ebb0a5

Bot/Chat A: Project A

Bot/Chat B: Project B

Additional Information

The two projects can use completely separate disk folders. The issue is not caused by shared project files.

The issue occurs because separate Grok Bot chats/bots are backed by the same underlying agent, causing durable memory and conversation history to be shared between otherwise separate projects.

Recurring status beats amplify the issue because each Project B beat refreshes Project B’s information in the shared agent context.

Ask

Scope durable memory and conversation history to the individual bot/chat or project, even when multiple separate bots use the same underlying agent.

There should be a hard context boundary so that activity, memory, status beats, and findings from one bot/project cannot be used to answer another bot/project.

Does this stop you from using Cursor?

No — but running two separate research projects through different bots that share the same agent is unsafe without manual isolation.

5

Hey @LLMTester, thanks for the detailed report.

What you are seeing is how a Grok Bot agent works today, rather than something broken in your setup!

Each bot has one conversation and one memory, and that single history spans both its 1:1 chat with you and every group it is a member of. Group turns are tagged internally so the bot knows which room it is speaking in, but it is expected to draw on everything it knows, including its 1:1 history, and its saved memories belong to the bot, not to a chat. That means two long running projects on the same bot will cross over, exactly as you describe.

The reliable way to keep Project X and Project Y separate right now is to give each project its own bot. If you want to stay on one bot, you can also ask it to keep Project Y notes in files under its project folder instead of saving them to memory, but that relies on the model following the instruction, so separate bots are the safer option.

Chat scoped memory is something the team is actively working toward!