This tool can recover deleted AI chats from your phone – how it works

Researchers at Georgia Tech have built an experimental tool that can pull deleted conversations with AI assistants back out of an Android phone. The work targets assistants that run locally on the device rather than in the cloud, and it was tested on a purpose-built fake assistant rather than a commercial chatbot. The tool is called ORISA, and the study's lead author is Ph.D. student Ken Xu.
The setup was simple. The team created a note-keeping assistant inside the ChatterUI Android app, loaded it with 100 fictional facts about a project, and then deleted that information. ORISA then searched the phone's working memory for traces of the deleted chat that could serve as context for a new session. It found enough usable material to spin up a working session containing 26 of the original 100 facts. That is more than the original assistant session had access to at once, because the remainder had fallen outside its active memory.
The mechanism the researchers describe centers on batch tokens and key-value caches retained in process memory. To unpack that: when a model runs, it processes text as tokens and stores intermediate attention state in a KV cache so it does not have to recompute earlier context on every step. Those caches are held in the app's memory while the session is alive. Deleting the chat clears the app's stored conversation but does not necessarily scrub every copy the runtime left behind. ORISA reconstructs token sequences from what survives and rebuilds session-level attention state, which is what lets it resume a conversation from fragments.
In practice, that means recovery is probabilistic and time-sensitive rather than a guaranteed undeleting feature. The researchers note that the longer you wait, the harder recovery becomes, because ordinary phone activity overwrites memory. A reader should treat the 26-of-100 figure as a result from one controlled experiment with a bespoke assistant, not a general success rate across apps.
The motivation is forensic. The team frames local AI as lacking centralized logging and enforcement, which cuts both ways. A user's chats stay off cloud servers, but there is also no audit trail an investigator can pull. The researchers point out that bad actors can fine-tune models for crimes and then delete the evidence, and that this has made it hard for forensic memory techniques to recover anything useful. ORISA is meant to close that gap; Xu said it can help establish what someone shared with an assistant and what help they received.
The privacy implication runs the other way for ordinary users. Local assistants are often chosen precisely because the session stays on the device, using device memory instead of a remote service. This work suggests that "deleted" may mean "removed from the app's view" rather than "gone from the hardware." A likely trade-off is that any tool built to recover this residue has to read raw process memory, which means it needs privileges most consumer apps do not get; this is closer to forensic software than to something a casual snoop runs. The researchers also note that a reimplementation or a leaked copy of ORISA could in theory be used to reconstruct someone's deleted queries from a phone they can access, though the same memory-overwrite limit would apply.
For developers building on-device assistants, the practical takeaway is that session teardown deserves attention. Clearing the visible chat history does not necessarily clear KV caches, token buffers, or whatever the runtime holds. Explicit cache eviction, minimizing how much context is retained, and treating memory as a data store with a retention policy are reasonable defaults. For product teams, the same reasoning applies to any feature where the pitch is privacy through locality: locality limits exposure to the cloud but does not by itself make data unrecoverable.
The broader advice that follows is unchanged: do not put sensitive information into a chatbot, even one running locally. What this research adds is a concrete reason, grounded in how inference runtimes actually store state.
Why it matters: Developers shipping local assistants now have evidence that deleted sessions can leave recoverable traces in process memory, which argues for deliberate cache clearing and tighter context retention rather than relying on history deletion. Security and forensics teams gain a research direction for pulling evidence from devices that keep no central logs. For end users, the practical change is modest but real: local does not mean ephemeral, and the window for recovery narrows only as new activity overwrites memory.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
This new tool proves it’s still best to limit the info you share with chatbots.