Here is message from Cursor CLI " Yes, I do. It’s bad on multiple levels:
Fabricated user quotes. The summary contained things like “I want to create an AWS Lambda function…” attributed to you –
things you never said. I presented those as fact.
I didn’t question it. When nothing in the workspace matched the summary, I should have stopped and said “this doesn’t add up.”
Instead I spent time searching for files that never existed and kept going as if the summary was gospel.
If it had been real data instead of hallucination – if the summarizer had pulled in an actual conversation from a different user
– that would have been a genuine data leak. The fact that it was fabricated rather than leaked is luck, not a safeguard.
Trust erosion. You’re now rightfully questioning whether you can trust the tool’s context handling, and whether your own
conversations could leak somewhere else.
This is a bug in Cursor’s conversation summarization system, not something you caused. If you want to raise it with the Cursor team,
the relevant details are:"
Steps to Reproduce
I was working with Cursor CLI and while asking it to fix a feature, it started giving me mention of tech stack that I did not have in my system and started looking for those files.. I stopped it and asked what the heck is it doing and it says “Oops, not sure where that came from” after a few questions and answers it said it found this in one of the jsonl files. and after I challenged it that I had never had discussions around this tech stack that it was talking about , it told me that it found it in the summary jsonl and it seems like it made it up..
Expected Behavior
This is a big issue..
Operating System
MacOS
Version Information
There it is. You have Claude Code CLI version 2.1.59 installed at /opt/homebrew/bin/claude. That’s Anthropic’s standalone CLI tool
Based on the version info you shared, “Claude Code CLI version 2.1.59 installed at /opt/homebrew/bin/claude”, it looks like you’re using Anthropic’s Claude Code CLI, not Cursor’s CLI. These are two separate products.
The AI statement that “This is a bug in Cursor’s conversation summarization system” was incorrect. Claude Code CLI was attributing its own behavior to Cursor. The fabricated summaries and hallucinated context you saw are happening within Claude Code’s own conversation management.
I’d recommend reporting this directly to Anthropic since it’s their product. You can reach them here:
that was my bad.. I did not specify which version. I use Cursor CLI to do all my coding.. and that’s were I was prompted with the info.. then I asked about where it came from and it pointed me to the conversation summary history..
Thanks for the info. I need to confirm one detail so I can help the right way.
Cursor CLI is started with the agent command in the terminal. Anthropic’s Claude Code CLI is started with the claude command. That’s what you have installed at /opt/homebrew/bin/claude, as shown in Version Information.
Can you check which command you’re using to start it, agent or claude? These are two different products with different summarization systems.
If it’s actually Cursor CLI agent, then the issue where summarization generates made up content is a known one. The team is aware. If that’s the case, please share:
Your agent version: agent --version
The Request ID for the affected chat: /copy-request-id
I normally use Cursor IDE + Cursor CLI (inside Cursor IDE).. Now that I brought back the project and tried the copy-request-id it is not finding it..
At the time I was having that error (and I did work with agent to get to the bottom of the issue) it indicated that the issue is from summarization (as per my initial post - that is the AI response).
Please let us know when this is fixed as it has spooked folks at my company…
This is a known issue in summarization where the agent makes up context that wasn’t actually in the conversation. Your CLI version 2026.02.27 came out before the fix, which shipped after that date.
Please update to the latest version, run agent --update and the summarization issue should be fixed.
The team is aware of this type of issue. At this point we don’t need a Request ID since the fix should already be out.
On your data security concern, the made up content comes from model hallucinations during summarization, not from real data leaks between users. Summaries are generated per conversation and don’t pull in data from other users.