ReRecall

ReRecall is a proof of concept for an open source implementation of Microsoft's Recall feature. It periodically takes screenshots of the user's desktop and stores them in a database. The user can then search through these screenshots using a text query.

THIS IS A PROOF OF CONCEPT, EXPECT A LOT OF PROBLEMS.

Requirements

You need a modern python version and a docker installation with GPU support, the docker-compose.yml file is configured for nvidia gpus. I have ran this on a Windows 10 with a GTX 1060. No part of the code is os-specific so it should work fine on other systems. Feel free to play around.

Setup

Clone the repository

git clone https://github.com/AbdBarho/ReRecall.git
cd ReRecall

Start Ollama

docker compose up

Open another terminal, and setup a virtual environment and install the requirements:

python -m venv .venv
source .venv/bin/activate
# or .\.venv\Scripts\Activate.ps1 on Windows Powershell
pip install -r requirements.txt

Now you can start the recording

python rerecall/record.py

By default it will capture screenshots of all monitors every 60 seconds, see rerecall/config.py for all options you can change.

Search

After recording some data, you can search through it using the search script. Ollama should be running in the background.

python rerecall/query.py -n 1 my search query

where -n 1 is the number of results you want to see, and my search query is the text you want to search for.

Notes

The current implementation is very primitive, it takes screenshots, uses ollama/llava-phi3:3.8b-mini-q4_0 to generate descriptions of the screenshots, and uses ollama/mxbai-embed-large to generate embedding for vector search, and stores them in a chroma db database.

For the search part, the embedding is done on the query text and matches against existing embeddings in the database.

You can change the models used for the description and embedding in rerecall/config.py.

There is definitely room for improvement here:

• There are plenty of hallucinations in the generated descriptions of screenshots, this could be a combination of the size the MLLM used to generate the descriptions (I use a very small model), or because the screenshots are very high in resolution.
• The search is very basic, it just matches the embeddings of the query text with the embeddings of the screenshots, a potential improvement could be to use the model to enrich the user query with more information before embedding it for search.
• I am fairly certain that Microsoft does not rely solely on screenshots as I do, but also captures of individual app windows, and also extracts meta information like window title & others, these could definitely improve the results.

Results

Cherry picked, ignore the vertical monitor.