Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarTechXplore AI→ original

Deleted AI Chat on Android Recovered From Memory: ORISA Pulled 26 of 100 Facts

A deleted chat with an on-device AI assistant on Android can leave traces in the phone's RAM. Georgia Tech researchers built ORISA, a tool that gathers these fragments into a working session. In the test, the assistant was given 100 invented facts and the chat was then deleted. The recovered copy named 26 facts, while the original session remembered only 14. Developers of on-device AI apps should therefore clear memory explicitly after a conversation is deleted.

AI-processed from TechXplore AI; edited by Hamidun News
Deleted AI Chat on Android Recovered From Memory: ORISA Pulled 26 of 100 Facts
Source: TechXplore AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTechXplore AI ↗
Editor-in-Chief
Key metric #1Tested 10 Android apps and 3 models in 60 configurations
Key metric #2The more memory is overwritten by other tasks, the harder it is to find traces
Key metric #3The work was presented at the IEEE Symposium on Security and Privacy 2026

You deleted your conversation with an AI assistant, and the interface shows an empty chat list. It looks like everything is gone. But researchers from Georgia Tech have shown that on Android phones running local models, this is not the case. Their ORISA tool finds fragments of already deleted conversations in the phone's RAM and assembles them into a working assistant session. In the experiment, the assistant was given 100 invented facts about a project. After the chat was deleted, ORISA recovered 26 of them, while the original session remembered only 14.

Why local AI does not mean leaving no trace

Local assistants run directly on the smartphone and do not send conversations to the cloud. That is why they are often considered more private: data does not go to someone else's server. But while the assistant is running, fragments of conversations sit in the phone's working memory, and deleting a chat in the app does not necessarily clear those memory regions.

The analogy is simple. You tossed a note into the trash, but the garbage has not been taken out yet. The interface already shows that the note is gone, but it is physically still there. ORISA essentially digs through that trash. It searches memory for pieces of deleted dialogues and reassembles them into a new session that you can talk to like a normal assistant.

How it was tested and what came out

For the demonstration, the authors built a fictional assistant called Project Note Keeper on top of the ChatterUI Android app. They "told" it 100 invented facts about a project and deleted the dialogue. ORISA recovered the information and built a new session from it. When asked to list the project notes, it named 26 facts.

The most interesting part is the comparison with the original. Before deletion, the original session could recall only 14 facts: the rest did not fit into its active context window, meaning the model no longer took them into account. The recovered copy retrieved more. This means the phone's memory retains not only what the model is using right now, but also what it has already "forgotten."

The scale of the test: 10 Android apps, three models, and 60 test configurations. In every case, the tool recovered information about conversations while it remained in the phone's memory. The paper is titled "Recovering and Rehosting Mobile Local LLM Conversations and Contexts via Memory Forensics" and was presented at the IEEE Symposium on Security and Privacy 2026. The authors are PhD students Ken Xu, David Oygenblik, Junze Zhang and Mingxuan Yao, research scientist Muhammad Ibrahim, and assistant professor Brendan Saltaformaggio.

Limitations and what to do about it

The method has a limit: the more memory is overwritten by other phone activity, the harder it is to find anything. This is not a way to pull up any conversation from months ago. In addition, the experiment was run in a lab with invented data, so it does not look like a mass threat to ordinary users yet.

There are two practical takeaways.

  • For forensics, this is a new source of evidence. According to Ken Xu, the tool helps establish what information a person gave an AI assistant and what help they received in return.
  • For privacy, this is a signal that "local" does not mean "leaves no trace." Anyone who gets access to a running device and knows how to analyze its memory could potentially see deleted conversations too.

Developers of mobile apps with local models should check what happens to context after a chat is deleted. Are buffers zeroed out, is the model state reset, are there any extra copies of the history during processing? Product teams that promise privacy through local execution should phrase those promises more carefully and should not claim complete erasure if it is not technically guaranteed.

Ordinary users should keep in mind that deleting a chat is not enough if it contained sensitive data. Closing the app and restarting the phone so memory gets cleared looks like a sensible step. But this is our own conclusion from the logic of the method, not something the authors tested directly.

Story

Key Questions Answered

Why local AI does not mean leaving no trace?

Local assistants run directly on the smartphone and do not send conversations to the cloud. That is why they are often considered more private: data does not go to someone else's server. But while the assistant is running, fragments of conversations sit in the phone's working memory, and deleting a chat in the app does not necessarily clear those memory regions. The analogy is simple. You tossed a note into the trash, but the garbage has not been taken out yet.

How it was tested and what came out?

For the demonstration, the authors built a fictional assistant called Project Note Keeper on top of the ChatterUI Android app. They "told" it 100 invented facts about a project and deleted the dialogue. ORISA recovered the information and built a new session from it. When asked to list the project notes, it named 26 facts. The most interesting part is the comparison with the original.

What are the key details on Limitations and what to do about it?

The method has a limit: the more memory is overwritten by other phone activity, the harder it is to find anything. This is not a way to pull up any conversation from months ago. In addition, the experiment was run in a lab with invented data, so it does not look like a mass threat to ordinary users yet. There are two practical takeaways. For forensics, this is a new source of evidence.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…