multi_object_memory_2025

Working Memory of Multi-Object Scenes in Primate Frontal Cortex

Introduction

This webpage is the entrypoint for the code and data supporting the paper “Working Memory of Multi-Object Scenes in Primate Frontal Cortex” (Watters et al., 2026). This paper studies the neural mechanisms of multi-object working memory in the primate brain. It involves neurophysiology data collected from monkeys trained on a working memory task and computational modeling to test hypotheses of neural mechanisms in that data.

This website is intended for users who want to replicated the results in our paper or want to use our data for your own research.

Here’s a video of our multi-object working memory task (played by a monkey whose gaze position is the green cross):

And here’s a video where you can listen to neurons recorded simultaneously with the task: videos_with_neurons/Perle_2022-05-31.mp4

Codebase Organization

Our codebase is written entirely in python, and consists of the following directories:

Usage

Getting Started

Navigate to our codebase: https://github.com/jazlab/multi_object_memory_2025

Download that repository by downloading it as a zip from GitHub or cloning using git commands.

Before running any code, we recommend using a virtual environment (e.g. using miniconda or pipenv) to manage dependencies. Once in a virtual environment with python version at least 3.10.0, install the necessary dependencies with:

pip install -r requirements.txt

Download and familiarizing Yourself With Our Data

Go to https://osf.io/vyw49 and download data_for_figures.zip. This is ~4.5 GB zipped and ~10.5 GB unzipped, so be sure you have enough storage space. Unzip the resulting file. It will contain folders behavior_processing, dandi_data, figures, modeling, and phys_processing. Move these folders all to the cache/ directory.

Now you can familiarize yourself with our data by doing two things:

  1. Visualize a trial by opening and running the Jupyter notebook ./task_visualization/demo_replay_trial.ipynb.
  2. Plot a neuron’s PSTH by opening and running the Jupyter notebook ./phys_processing/psths/demo.ipynb.

To further familiarize yourself with our data, you can also look at PSTHs and summary plots for all of our recorded units from our OSF dataset. In the phys_processing folder of the cache you downloaded and unzipped, you will see the folders psths and unit_summary_plots. Inside psths you can find PSTH plots for all of our units labeled good (single neurons) by our spike sorting pipeline. Inside unit_summary_plots you can find PSTH plots for all of our recorded units. We recommend looking through some of the PSTHs to both the diversity of responses across units and get an intuition for common patterns of modulation. Our best sessions are Perle/2022-05-31 and Perle/2022-06-01, so we recommend starting there.

Reproducing Figures

The code to generate each of the figures in our paper is in ./figures/. Each directory there corresponds to one figure, with Jupyter notebooks that generate plots for each panel in the figure. After you have downloaded our data and propagated the cache directory, each of these notebooks can be run.

Full Data Processing Pipeline

We have open-sourced all of our data, including behavior, task, raw physiology, and spike-sorted physiology, on DANDI as dataset 000620. You can download that data using the script ./download_dandi_data.py (see documentation at the top of that file for instructions). However, note that the raw physiology data is several terabytes, so we recommend only downloading the behavior/task and spike-sorted physiology data.

To replicate our data processing pipeline from the DANDI data files, you must first convert the raw task/behavior files to dataframes by navigating to ./behavior_processing and running run_cache_data.py there.

To process the spike-sorted DANDI physiology files, for each unit you must extract spike times per trial and compute AUC selectivity metrics. See the documentation and scripts in ./phys_processing for how to do that. These scripts take several hours to run on a laptop, so we do not recommend running them. We’ve provided the results of these scripts in /cache/, so you do not have to re-run them to reproduce our figures.

If you want to spike-sort our raw data, we have provided the raw physiology files on DANDI. We spike-sorted these by doing drift correction, running Kilosort 2.5, applying custom automated quality labeling pipeline, and manually curating the results. Our spike-sorting pipeline is not included in this package, but we recommend looking at the SpikeInterface toolkit if you’d like to spike-sort yourself.

Contact and Support

Please email Nick Watters at nwatters@mit.edu with questions and feedback. If you plan to use our data for your own research, we encourage you to email Nick Watters (nwatters@mit.edu) or Mehrdad Jazayeri (mjaz@mit.edu) so we can help you determine whether our data is useful for your scientific goals and discuss any limitations or challenges that might arise.

Reference

If you use our data or code (or a derivative of our data or code) in your work, please cite it as follows:

@article {Watters2026.01.27.702062,
	author = {Watters, Nicholas and Gabel, John and Tenenbaum, Joshua and Jazayeri, Mehrdad},
	title = {Working Memory of Multi-Object Scenes in Primate Frontal Cortex},
	elocation-id = {2026.01.27.702062},
	year = {2026},
	doi = {10.64898/2026.01.27.702062},
	URL = {https://www.biorxiv.org/content/early/2026/01/30/2026.01.27.702062},
	journal = {bioRxiv}
}
This site is open source. Improve this page.