Researchers recover deleted Android AI chats and rebuild working sessions

Deleting a chat with an AI assistant running on an Android phone may leave enough information to rebuild a working session. Georgia Tech researchers developed ORISA to recover those remnants from phone memory—a finding that offers digital investigators another way to retrieve information while raising privacy concerns for users of local AI.

A deleted chat can outlast deletion

The research concerns local assistants: AI systems that run directly on a phone rather than sending conversations to a cloud service. While an assistant is running, pieces of a conversation can remain in the phone’s working memory even after the conversation has been deleted.

ORISA searches that memory for conversation fragments. Its task goes beyond retrieving isolated pieces of text: the researchers can use recovered information to reconstruct a usable AI session. The distinction is between finding traces of a deleted exchange and making those traces available to an assistant again.

The team tested the tool with 10 Android apps and three AI models across 60 configurations. It recovered conversation information remaining in phone memory across those tests. That result describes recovery of surviving information, not a guarantee that every deleted conversation can be restored in full.

The restored assistant recalled more

An experiment using the ChatterUI Android app illustrates the difference between information an assistant can actively use and information still recoverable from the device. Researchers created a fictional assistant called Project Note Keeper, supplied it with 100 fictional facts about a project, and then deleted the conversation.

ORISA recovered information from the deleted exchange and used it to create a new working session. Asked to list its project notes, the restored assistant returned 26 facts. The original session could access only 14 because the remaining information had fallen outside its active memory.

The restored assistant did not return all 100 supplied facts. But it did retrieve information the original assistant had already stopped actively using. The experiment therefore demonstrates more than recovery after deletion: it shows that an assistant’s current access to information need not mark the limit of what remains recoverable.

Evidence for investigators, exposure for users

For digital investigators, that capability could help reconstruct both sides of someone’s interaction with an assistant. Lead author Ken Xu, a Georgia Tech PhD student, said the tool could help establish what information a person shared with an AI assistant and what assistance they received.

The privacy concern is the other side of that same recovery capability: a user can delete a local conversation without eliminating every trace. Keeping the assistant’s processing on the phone, rather than sending the exchange to a cloud service, does not by itself resolve the memory-remnant problem demonstrated here.

Recovery also has a practical limit. The researchers say information becomes harder to retrieve as other phone activity overwrites the relevant memory. ORISA’s results depend on conversation information still being present, even if the chat has disappeared or the assistant can no longer actively use all of it.

The study, Recovering and Rehosting Mobile Local LLM Conversations and Contexts via Memory Forensics, was presented at the 2026 IEEE Symposium on Security and Privacy. The work was conducted in Georgia Tech’s Cyber Forensics Innovation Laboratory, which connects the university’s cybersecurity and electrical and computer engineering schools.