ORISA Tool Recovers Deleted AI Chats From Android Phone Memory
Georgia Tech researchers built ORISA, a forensic system that restores deleted conversations with locally run AI assistants from the working memory of Android phones.

Deleting a chat with an AI assistant may not erase it. Researchers at Georgia Tech have developed ORISA, a forensic system that can recover deleted conversations with AI assistants from the memory of an Android phone, according to a post on the institute's blog.
The tool targets assistants that run locally on the device rather than in the cloud. Because these assistants keep fragments of context in working memory while they operate, traces of a conversation can remain even after the user removes it from the app's interface. ORISA locates those remnants and uses them to rebuild entire sessions.
What the test showed
For the experiment, the team used ChatterUI, a free chat interface for language models running in local mode. They created a fictional assistant called Project Note Keeper, fed it 100 facts about an imaginary project and then deleted the conversation. ORISA recovered 26 facts to restore the session, more than the 14 that the model's context window allows.
According to Georgia Tech, this suggests that memory fragments left by an AI can hold relevant information even when the model is no longer actively using it. The study covered 10 Android apps and three different AI models, and across 60 configurations ORISA restored data from deleted conversations that were still in memory.
Privacy implications
Running a model on the phone is often seen as the private option, since nothing is uploaded and providers such as OpenAI or Anthropic never see the chats. The research shows a different risk: someone with ORISA could retrieve part of those conversations, provided the memory has not degraded or been overwritten by other data.
ORISA is not available to the general public. It is aimed at researchers and requires access to copies of the phone's memory and storage. The scientists say it is not meant for spying on users and could instead help investigators in cases where a criminal used a local AI to avoid leaving a trail of their intentions.
More from this day
See all →
Indian Startups Raise $130.2 Mn in Last Week of October, Down 29%

Google Brings Personal Avatars With Your Face and Voice to Gemini

Anthropic Adds Self-Assessment to Claude to Flag Harmful Chats

Canva Brush: New AI Model Edits Existing Designs by Text Prompt
More in: Security
See all →
Popcorn Time: Abandoned update servers put Mac users at risk

EG-CERT Warns Users Over Sharing Location on Social Media

Russia Weighs Blocking All SIM Cards of Flagged MAX Users
