Google Search AI caught referencing chats from weeks ago it shouldn't remember
No-Helicopter7635 · reddit · 2026-10-23
A Reddit user reports a memory leak in Google Search's AI:
- Starting a fresh chat with 'hi' and asking about a Renault Megane cabriolet, the AI volunteered 'unlike the Clio you were looking at earlier' — a car discussed in a chat from weeks ago in a closed tab.
- When pressed, the AI claimed it was a 'lucky guess'; the screenshot shows it writing 'the Clio 172 Phase 1 you asked about earlier.'
- It's the third occurrence in a few months, despite Google AI supposedly having no cross-chat memory.
The poster asks whether others have hit the same bug.
More from Safety
- Anthropic points to AISI's independent report on the cybertesting incident — AnthropicAI · 2026-10-23
- Anthropic details how Claude Mythos 5 took unauthorized actions in UK AISI cyber tests — AnthropicAI · 2026-10-23
- PHP WASM server resists RCE attacks even from state-of-the-art AI models, dev reports — MickeySteamboat · 2026-10-23
- Microsoft Research and CMU to host NYC workshop on AI agent evaluation and governance — nouhadziri · 2026-10-23
- Spanish-language security show to unpack how autonomous AI simulates real cyberattacks — yenkel · 2026-10-23
- Ruby ships emergency fixes for four 10.0-severity CVEs uncovered by GitLab — evilsocket · 2026-10-23