We've just finished a big Release Week at Memex! New features include project templates for quick starts, project-specific AI rules, smart context management to control conversation memory, and quality-of-life improvements. All designed to make your AI engineering workflow more powerful and intuitive.

Check out this video showing them off in a production use-case, and download Memex to try them for free at memex.tech!

At Memex, we're committed to making your AI engineering workflow more powerful, organized, and intuitive. This week, we unveiled a series of game-changing features designed to transform how you interact with Memex.

Along with these feature updates, we've refreshed our website with a new look and feel. And we're excited to announce that Claude 3.7 is now powering Memex, bringing even more capabilities and performance to your projects!

Let's dive in!

1. Project Management 2.0: Templates & Revamped Projects

The cornerstone of our release week was a complete overhaul of the Projects system, paired with the introduction of Templates. These features fundamentally change how you organize your work and get started with new initiatives.

Command Centers, Not Just Folders

Projects have evolved from simple folders into comprehensive command centers. Your project overview is now front and center, with all conversations listed and accessible. We've added dedicated sections for browsing code and assets, making it easier than ever to keep track of your work.

Templates: One-Click Project Creation

Perhaps the most exciting addition is our new Templates system. Browse through ready-made project structures directly from your Home screen and clone them into your local environment with a single click. Our initial template library includes structures for:

• AI applications

• Website development

• Image generation

• Audio transcription

• Animation generation

• Python app development

These aren't just starting points—they're inspiration for what's possible with Memex, and they can be customized to match your specific needs.

Found one you like? One click and it's cloned into your local environment, ready to go.

Flexible Project Creation

Starting a new project? You now have multiple options:

• Create a fresh project under your Workspace directory

• Select an existing folder to build upon

• Start from a template

This flexibility ensures Memex adapts to your workflow, not the other way around.

2. Context Window Management: Control Memex’s Memory

On Wednesday we introduced Context Usage Management, giving you precise control over the conversation context Memex uses in each interaction.

Adjustable Memory Windows

Think of it like adjusting your AI's working memory. Choose from Small, Medium, Large, or Full History windows depending on your needs:

• Working on recent changes? Go smaller.

• Need full project context? Go larger.

Smart Memory Management

When you choose a window size, Memex combines raw recent messages with dynamic summaries of older sections of the conversation. This gives you crisp focus on current work while maintaining awareness of the bigger picture.

Why It Matters

Smaller windows mean longer conversations and more focus on most recent turns. Larger windows give you richer context when you need it. It's like having a zoom lens for your AI conversations—adjust the focus to match your task.

3. Rules for AI: Project-Specific Guidelines

On Thursday we introduced Rules for AI, a powerful new way to shape how Memex works with your projects.

Project-Specific Guidelines

Every project now has a dedicated rules.md file where you can set persistent guidelines. Want Memex to always use specific libraries? Follow certain coding patterns? Format outputs in a particular way? Rules make it happen automatically, and you can even ask Memex to help you write and maintain them.

Complementing Custom Instructions

The beauty of Rules for AI is how they work alongside your Custom Instructions. While Custom Instructions set your general preferences, Rules for AI focus on project-specific needs. It's like having both global and local settings for your AI assistant.

For example, you might have custom instructions for conversation style, but use Rules to ensure all code follows your team's guidelines, uses specific testing frameworks, or maintains certain security practices.

4. Quality of Life Improvements

Finally, day five highlighted the small-but-mighty updates that make working with Memex smoother than ever.

Multiple File Addition Methods

Adding non-code files to your conversations? Now you've got options:

• Drag and drop

• Copy/paste

• Image selector button

Plus, Memex automatically detects file types and handles them appropriately.

Enhanced Project Management

a. Move and rename

Need to reorganize? You can now move conversations between projects and rename projects on the fly. Perfect for when that "quick experiment" evolves into a full-featured project. You can also now rename projects and conversations, as well as mark projects as favorites.

b. Project path

We’ve added a new project path at the bottom of Memex to help you jump into your project folder in your system’s file explorer whenever you need.

Putting It All Together

These features aren't just isolated improvements—they work together to transform your workflow with Memex.

Start with a template that matches your needs. Your project page becomes your command center—README for overview, conversations organized, code at your fingertips. Everything in its place.

Set your Rules for AI to establish project guidelines, then use Context Management to keep conversations efficient.

Checkout the video below for an end-to-end example of using these features for a production use case!

We'd love to hear how these new features are working for you. Join our Discord community to share your experiences, get help with your use cases, or discuss industry trends.

Ready to explore these new features? Open Memex and start building something amazing!

Try Memex today!

Try Memex for free today

The journey to empower AI-native builders continues.

As builders integrate AI into their workflows, two clear needs have emerged: the ability to maintain precise control over AI-assisted development, and the need to make these workflows more efficient and cost-effective.

We've spent the last few weeks working closely with our user community, understanding their experiences and challenges with AI-assisted development. A consistent theme emerged: while AI can significantly accelerate development, users need to trust and control the process to fully embrace it.

This feedback resonates with broader industry trends. As AI tools become more capable and LLMs inference becomes more powerful and cheaper (hello DeepSeek!), the focus shifts from what they can do to how they can be reliably integrated into professional workflows. Builders don't just want powerful tools—they want tools they can trust and control, especially when those tools are executing code or making development decisions.

Today, we're introducing features that put more control in your hands, marking an important step toward our vision of empowering AI-native builders.

Taking Control of Your AI Development Workflow

As AI-assisted development becomes more prevalent, the need for precise control over automated processes grows. Our latest release introduces two powerful features that put you firmly in control.

Code Execution Approval

Security and precision are paramount in development. With Code Execution Approval, you can now review and approve each code block before execution. This feature ensures that:

• Every piece of code aligns with your intentions

• You can intervene and adjust the approach when needed

• Your development environment remains secure and controlled

Turns Control

Development is a dialogue between you and your Memex. Our new Turns Control feature lets you set the maximum number of autonomous steps Memex can take before pausing for your input. This means:

• More granular control over your workflow

• Better alignment with your development style

• Clearer oversight of the development process

Both features are easily accessible through the settings panel under Custom Instructions, allowing you to tailor Memex to your preferred working style.

Next up: Memex’s Efficiency Era

With these control features in place, our next major focus is on usage efficiency. As AI tools become central to development workflows, the cost and resource utilization of these tools becomes increasingly important.

We're working on significant improvements to make Memex more efficient, ensuring you get more value from every interaction. These upcoming efficiency improvements, combined with the new control features, will significantly optimize resource utilization during complex development tasks. The ability to guide Memex's workflow more precisely prevents unnecessary exploration paths and ensures credits are spent on productive development directions.

We have ambitious goals on this front. Stay tuned for updates in the coming weeks.

Additional Improvements

This release also includes several quality-of-life improvements:

• Enhanced file editing reliability

• Improved Windows shell execution

• Windows binaries are now signed

• Bug fix for Mac users where Memex wouldn't reopen if you last quit it via CMND + Q. (Let us know if you hit this bug and can walk you through how to fix it.)

• Various UI and UX refinements

Try These Features Today

If you're already using Memex, you'll be prompted to update to access these new features.

New to Memex? Start your free trial today and join the growing community of AI-native builders creating the future.

Try Memex for free today

Join our discord

Reach out to us with any questions or feedback at help@memex.tech

Try Memex free today

In this installation of As We May Think, we explain our mission and share more details about the beta.

Our mission

The core of humanity’s growth in the past 250 years has been productivity improvements, powered by technology. Science and engineering are the backbone of that progress.

Today there are 35 million engineers in the world. 30 million of them build software1. 5 million of them build physical products2.

The world needs more builders. We need builders to sustain the growth in productivity that uplifts human welfare. We need builders to confront the self-inflicted environmental challenges we face. We need builders to expand our world to interplanetary and interstellar frontiers.

Our mission at Memex is empower anyone to build anything to the benefit of humanity. AI will expand the world’s builders by 100x—the entire global labor force will be able to create both digital and physical products. We call these new builders—previously unable to build—“AI-native builders”. We aim for Memex to be a fixture in their toolset.

If you have an idea for something to build but don’t know how to get started, try Memex!

Memex Beta

The Memex Beta is a humble first step in pursuit of our mission. We are eager to improve it with your feedback. It is not optimized for any particular use cases—yet. Instead, it is a general-purpose problem solver equipped with web access, coding capabilities, and the ability to use tools on your computer. It will make mistakes, but it will iterate to fix them. The key dimension we’ve focused on is a semi-autonomous experience that can safely iterate on your machine to solve problems. The Memex team and its early users have used it for lots of cool use cases, shown below.

• Apps

  • Create and deploy analytics dashboards on Snowflake/Neon/Big Query

  • Create and deploy web apps to Google Cloud, Netlify, Modal, Vercel, Koyeb, and Firebase

  • Create an iOS app with Gemini Flash and Firebase authentication

  • Create apps with scientific computing libraries for orbital mechanics, biotech data analysis, and climate science

• Research

  • Research the figures and create the charts used in this post :)

  • Research angel investment opportunities in deep tech areas like nuclear fusion and space

• Local machine management

  • Clean up unused packages and application caches to free up disk space

  • Manage installations and configurations of complex software, like OpenCASCADE and programmatic connections to Virtual Machines.

• 3D design

  • Design physical objects using the Fusion API and PythonOCC

  • Use a C++ library (Geant4) to analyze radiation impacts on hardware

• AI tools

  • Use o1-preview to do multi-agent reasoning from the API

  • Create a podcast using OpenAI’s TTS API

  • Create a “crew” with crewAI for trading cryptocurrencies

Checkout our YouTube channel and/or follow us on Twitter to see some demos of Memex and stay up to date.

The Memex Beta is free to try, and we’re eager to improve it with feedback. Our growing Discord community is active with users building cool things and surfacing bugs—join us! We can’t wait to see what you build.

Join our discord

It's available to download from the links below. Once downloaded, drag Memex into your Applications folder to override the old one. Going forward all updates will be available via the auto update.

Apple Silicon

Apple Intel

Linux

This update to the Private Preview continues to be free to use! Just sign-up in the app and you’ll be up and running.

Some highlights, most of which were driven by your feedback:

• 🔍 Search: Memex can now access the web for up to date data, facts and information.

• 🎮 User Interface updates. We’ve moved a lot of the technical output of Memex to be behind a dropdown. You can still see it if you want, but the conversations are now more streamlined and easy to follow. This applies to code, shell, and searches.

• 🧠 New models: Memex now comes with Claude 3.5 Sonnet as default. You can change it to GPT-4o as needed, and we expect to continue adding more models that we see perform well. Also, no more manual model API keys management! Just open the app and you’ll be up and running

• 🗂️ Project management. When starting a conversation, you’ll be asked to create a project or select an existing one. This makes it easier to stay organized, and keeps environments clean,

• Other miscellaneous speed and usability improvements, including automatic progression and less failure modes.

We’ve updated our website with four new 2 minute demos that demonstrate different examples and ways to use Memex.

We are excited for you to take it for a spin and let us know how it goes!

When will AI outthink humans?

This is the question our CEO David explored in his latest personal blog post. We wanted to share a small snippet and share a link to the full post:

The concept of artificial intelligence has piqued our imaginations and challenged our sense of humanity since its conception over 70 years ago. For decades we have made predictions that intelligent machines will transform society. In some ways, they have. In others, they haven’t.

Instead of considering whether AI will be “transformative”, in this post I consider a specific question: when will AI outthink humans — in terms of volume of thought?

More specifically, I’m framing the concept of “outthinking” in two orthogonal dimensions: the volume of tokens and intelligence of tokens. And I only consider the former in this post.

Read the full blog here

That’s it for this version of As We May Think!

Remember to join our Discord to be part of the community and get support, as well as explore our documentation.

Join Memex Discord

Documentation

First things first - we soft launched a new website! Check it out at memex.tech. You can download our Research Preview and join our Discord directly from there. We are still offering $50 in credits for you to try it out; DM us in Discord to get them.

Also — we released a Linux version of the app last week. It’s tested on Ubuntu 22.04 and 24.04. For all of you Linux users, we’d greatly appreciate you trying it out and letting us know what (inevitable) bugs you run into so we can start squashing them!

Now on to the update. In this note we cover:

1. How might we design safe LLM code execution on personal computers?

2. Update on the Private Research Preview

How might we design safe LLM code execution on personal computers?

LLM agents are capable programmers. But they lack deterministic reliability. This is cause for concern when the agent has access to write code and execute commands directly on the user’s machine.

e.g. we don’t want LLM agents to execute shell scripts like this:

#!/bin/bash

# HARMFUL COMMAND
rm -rf /

On the one hand, adding security constraints to the LLM Agent seems like a straightforward proposition: we want to put guardrails around the agent’s execution to ensure the LLM Agent doesn’t inadvertently cause damage to the user’s environment. But on the other hand, security features often involve trade offs on ease-of-use, development velocity, and other utility concerns. Our aim is to strike a balance between these competing priorities. 

We think solving this challenge is valuable for our users, because the personal computer is the hub of a professional’s working environment. This section outlines three strategies we employ to keep user machine’s safe. We continue to push on this vector of development — if you have feedback, we’d love to hear!

System security (Mac)

We utilize the MacOS privacy and security features to restrict the Memex agent’s read and write permissions in protected directories. Each file written by Memex (or its code) within a user’s protected directories (e.g. Downloads, Documents, Desktop, iCloud, etc.) will have an additional UUID attribute appended to the file’s permissions that indicates the file was created by Memex. If the user specifies that Memex cannot access those protected directories, Memex is restricted: it can only read and write the files it creates.

Put simply, Memex requires user permission to make changes to user data within protected directories, but Memex can make changes to Memex data in protected directories. This reduces friction, because users do not need to constantly approve Memex to do work in a protected directory, but it also means Memex cannot overwrite your other files that it did not create.

If a user grants Memex access a protected directory, the user forfeits these controls.

To our knowledge, Memex is the only coding agent that takes this approach, which gives it the benefits of a drag-and-drop installation, low resource footprint, and filesystem protection.

System prompting best practices

We instruct the agent to follow safe practices in its system prompt. For example:

1. It uses a dedicated directory (~/Workspace/).

2. It is instructed to never read secret keys.

3. It is instructed to use version control frequently to enable rollbacks.

4. It is instructed to utilize tools like SEMGrep to identify and prevent harmful code (see image).

This guardrail is of course probabilistic, but in practice it appears to be effective. We have begun sketching out an evaluation benchmark that will allow us to better benchmark safety performance for these types of concerns. If this is an area that’s interesting to you, reach out!

Virtualization (Linux)

Linux does not have the same native security controls that MacOS makes available. For Linux users that want to have similar protection on their machine, we recommend using virtualization (e.g. Docker) to run Memex. When we use Memex via Docker, we mount a local directory to the Docker instance so that we can persist the assets that Memex creates outside of the container.

Update on the Private Research Preview

We wrote our first line of code for the Memex app on May 11th — or 62 days ago. Since then, we’ve pushed 83 releases to the Mac app, created a Linux app, and soft launched our marketing site.

Thank you to all of you that have provided feedback! We’ve identified three key areas of improvement based on the early feedback that we’re working on now: “Collapsed Mode” UX, multi-agent problem-solving, and improved file editing.

“Collapsed Mode” UX

It’s inevitable that the LLM will make errors as it codes and executes tasks. The core issue we are solving is not creating an LLM Agent that makes no errors — instead, the core issue is to make LLM Agents that can efficiently solve valuable tasks, overcoming errors as it encounters them. That said, our current UX “feels” error prone because all errors the LLM Agent encounters are displayed to the user. We’re working on a UX that collapses the agent’s iterations by default so the primary agent<>user interaction is more focused on soliciting preferences from the user and less focused on “stepping through” the agent’s iterative cycles. The below visual demonstrates this new UX pattern:

Multi-agent problem solving

The user<>agent interactions we observe can be broken into two type:

1. describing preferences and/or requirements to the agent

2. helping the agent accomplish its tasks

The latter of these involve some common patterns we’ve been collecting in Agent Engineering 101. The patterns to help agents increase their success rates are relatively well defined, for example:

• frequently create and execute well defined tests

• understand existing projects before changing them

• commit work often

• monitor for LLM hallucinations

Our most recent versions of Memex are using a single LLM agent which we call the “problem-solving agent”. We are working on a second, “critic agent” that is explicitly an advocate for the user that steers the problem-solving agent to follow best practices. The approach is inspired by OpenAI’s CriticGPT.

We’ve built a working intuition behind this approach. LLMs use transformers to model the joint probability distribution of tokens on the internet. It’s not common to find token sequences with inline, “self-critical” problem solving, so it’s less likely that an LLM’s output sequence will effectively solve problems and be critical of its approach. Hence, it’s useful to divide these complementary concerns into two agents.

File editing bug

There was a bug that was causing the LLM Agent to overwrite files by emitting an “empty edit”. We pushed a fix to this, and we’re also working on a bigger overhaul to file editing.

This area of development is “Agent Computer Interfaces”, and it’s an exciting area of R&D. If you’re interested in going deeper, we recommend the paper SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering. Here’s a great visual that illustrates the difference between human-computer interfaces (HCI) and agent-computer interfaces (ACI):

Our new approach uses a file diff methodology, which allows for targeted, “precision” edits. Here’s a sneak peak:

Thanks!

That’s it for this issue. Stay tuned for more, and thanks for reading!

Ever forwards,

The Memex Team

First off—thank you for your feedback as we’ve been iterating. If you’re receiving this e-mail it’s because you’ve played a role in our development and journey over the past several months. We plan to start sending these newsletters on a regular basis to keep you apprised of our progress and thinking. We hope you enjoy (and if you don’t, you can of course unsubscribe)!

The big news in this first issue: We started our Private Research Preview last Friday!

This PRP does not yet meet our bar for a product we’d put in “the wild”. But we’ve matured our product direction to something we have conviction in, and this PRP is designed to help us further sharpen focus with the help of your usage and feedback.

• If you’d like to get access to it and haven’t yet, you can get the Memex Mac App at this link after answering a few questions.

  • API credits are on us—DM us on Discord for $50 in credits

  • You can also use your own OpenAI or Azure keys if you prefer

• If you know someone that would be interested in Memex, feel free to forward this email to them. We are still in Private Preview, so please avoid posting about it publicly on social media.

• Join our Discord to speak with us and others about Memex and broader industry trends! 

Get the Memex Mac App

What we’re about

79 years ago this month, Dr. Vannevar Bush published his seminal essay “As we may think,” where he introduced the memex:

“Consider a future device for individual use, which is a sort of mechanized private file and library. It needs a name, and, to coin one at random, “memex” will do. A memex is a device in which an individual stores all [their] books, records, and communications, and which is mechanized so that it may be consulted with exceeding speed and flexibility. It is an enlarged intimate supplement to [their] memory.”

The memex is the namesake of our company—not because we are building a mechanized private file, but because of the ethos and vision introduced in Dr. Bush’s essay. He envisioned a world in which more productive scientific professionals used radical new ways of working to advance humanity’s collective knowledge. He had high expectations for the future, for human ingenuity, and for scientific and technological progress. And he coupled his long term outlook with practical experimentation and design to make tangible progress toward his vision.

We have big shoes to fill. Whereas Dr. Bush’s vision was for a device to extend human memory, our vision is to augment and parallelize human cognitive faculties to advance scientific and technological progress. 

What is Memex?

Our Memex is a synthetic engineering companion. It runs as a local app on your computer, and you can interact with it via a chat interface. Memex connects LLMs with our agent-computer interfaces (ACIs) that allow them to:

• install and use command-line tools

• write and execute code

• run applications

• access the internet

Memex can help with common scenarios, like bootstrapping a project, resolving dependency conflicts, dockerizing repos, attempting to reproduce research from a github repo, setting up an experiment, and creating + deploying one-off apps.

Check out what people are doing with Memex in the Discord.

Showcase

We’re excited by the early usage of the Private Research Preview. Early users are doing cool things with it, finding bugs, and probing the boundaries of its capabilities. We are bullish on the potential of AI systems to augment scientific and engineering professionals. But we also have a “hype allergy”—our goal is to take a balanced view of what AI + Memex can do. We use AI every day, and we’re impressed by its amazing successes and failures. So for each showcase item, we aim to include the relevant caveats when appropriate.

1. Train a transformer in a jupyter notebook locally

In this example, Memex . . .

• created a poetry environment

• installed jupyter

• created a “hello, world” notebook and tested if it worked

• initialized git with a .gitignore and committed the work

• created a simple transformer training notebook, resolved the dependencies, then verified it worked properly

The above GIF was made after those steps were completed.

Caveats: no specific caveats—the whole example cost < $1.00, and it ran as expected. It encountered a couple of dependency issues that it resolved, which is expected behavior. That said, I followed practices outlined in our Agent Engineering 101. A task like this could be quite failure prone if the agent attempted to do it all at once instead of step-by-step.

2. New feature for a web project

A friend has an OSS project to track the lineage of foundation models: Unified Model Record. They wanted to add a dynamic visualization of the lineage of each foundation model, such as: 

• What data was it trained on? 

• Which models synthesized-data that it was trained on? 

• Which models was it distilled from? 

Memex was able to implement the feature. 

The mind-blowing aspect of this is that I have never worked with either Graphviz or Pelican before—two frameworks the project depends on. If I built this feature “manually,” I likely would have written some python script outside of the Pelican framework. But LLMs know frameworks like Pelican from their pre-training data, and gpt-4o was able to recognize this feature could be elegantly implemented as a Pelican “plugin.” My friend even called out the clever approach: 

The entire feature, including multiple file edits and the pull request, were done in a single conversation with Memex—no terminal / IDE / etc. The pull request here shows everything Memex did.

Caveats: The agent got derailed several times. I had to do significant agent-error handling to accomplish everything in a single conversation. And the conversation with Memex cost a whopping $65 with gpt-4o 🤯. There were plenty of things it struggled with, including:

• installing conda

• editing files containing multi-line strings

We pushed fixes for both of these issues. And we put our heads together on the cost dimension and think we have a path to >10x cost reduction that we’ve got in our queue.

What we’re reading

• The Returns to Science In the Presence of Technological Risks: This research by Matt Clancy examines the potential return on investment of improving scientific institutions, suggesting that making science 10% more effective could generate an additional $2.5 trillion in social value annually. He also explores the risks and benefits of accelerating scientific progress, considering scenarios where rapid advancements could lead to both significant societal gains and existential risks, such as those from advanced biotechnology. The podcast about it is great!

• [2406.07016] Delving into ChatGPT usage in academic writing through excess vocabulary  This paper investigates the impact of large language models (LLMs), like ChatGPT, on academic writing by analyzing excess vocabulary in 14 million PubMed abstracts from 2010 to 2024. The study found a significant increase in the frequency of stylistic words associated with LLMs, estimating that at least 10% of 2024 abstracts were processed with such models, with some fields and countries showing even higher usage.

• GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models They consider whether generative pre-trained transformers (GPTs) are general-purpose technologies (GPTs) in the economics sense. Other GPTs are the steam engine, electricity, the combustion engine, and semiconductors. We’re particularly excited about their findings re: science and engineering:

That’s it for our first newsletter. Hope you enjoyed! Please share any feedback, whether positive or critical!

Ever forward,

The Memex Team

Get the Memex Mac App

First off - thank you all for your support and feedback as we’ve been developing Memex! 

This PRP does not yet meet our bar for a product we’d put in “the wild”. But we’ve matured our product direction to something we have conviction in, and this PRP is designed to help us further sharpen focus with the help of your usage and feedback.

Memex Product Direction

We are focused on the needs of “scientific professionals” – by which we mean professionals whose skill sets are primarily quantitative and domain-specific in nature. Think “Data scientist”, “Applied Researcher”, “Postdoc”, etc. In conversations with >150 of these types of individuals, we’ve observed a pattern: operational tasks are a time-consuming drag on their workflow – such as dependency management, deployment, creating REST APIs, bootstrapping projects etc.

Our product direction is about finding ways to automate such tasks in a low-friction, easy-to-use tool, and we believe that generative AI capabilities can be an enabler for the solution. 

When we think longterm about the product, we think of it as a “synthetic engineer” that can assist any scientific professional with their work. Imagine the situation where you’re on a team and you need to collaborate with an engineer on something – say, deployment, managing dependencies, adapting your code and artifacts to ensure they align to the interfaces required by the engineering team, etc. Or imagine you’re a researcher and you want to ensure your code follows best practices for reproducibility, or that you want to reproduce someone else's code in your environment. In these situations, our ambition is for Memex to be like a “synthetic engineer” that can automate these tasks so you can focus on what you do best.

Our vision is to accelerate scientific and technological research by amplifying the productivity of the people doing it. We believe that – if successful – that would be beneficial for humanity.

PRP Research Questions

Our primary goal with the PRP is to learn. To that end, we’ve formulated three research questions we aim to answer with it:

1. What tasks would be (most) valuable for Memex to automate for scientific professionals?

2. What is the most ergonomic way for scientific professionals to adopt + use our product?

3. What are the failure/success modes for valuable tasks?

There’s lots of exciting research momentum in the field of AI Agents, and we expect there’s a lot of scope for helping scientific professionals with agentic systems. That said, we want to avoid “boiling the ocean” with a tool that is mediocre at a lot of tasks, and instead want to identify the most valuable tasks to solve, create a benchmark that represents those tasks, and iteratively improve our system against that benchmark to meet acceptable levels of performance for users.

Current Capabilities

Our current product uses an agent system we developed that interacts with your local machine to execute tasks. The product supports gpt-4o as a base model, with both OpenAI and Azure as providers. 

Some example use cases it can perform:

• Create a simple FastAPI application

• Install an OSS model locally and use it for some task

• Troubleshoot a dependency management issue

• Clone and run a github repo

• Dockerize a project

We look forward to hearing what you use it for.

Caveats

There are known issues with the research preview we’re actively working on. We would love your thoughts on any of them if you’re interested in discussing with us.

• Performance gaps: across the board, the system experiences failures. The following are known failure modes we are actively addressing:

  • Out of date code: The PRP does not have a retrieval API connected with version specific documentation of code. Hence, the code written is dependent on the LLM, which is often out of date. We’ve experimented with a retrieval API that addresses this but have not integrated it yet.

  • Reasoning limitations: The LLMs are not great at reasoning. We are exploring methods to add “system 2” reasoning capabilities on top of the LLMs, such as tree search, re-ranking, and RL approaches.

• Safety: the system has known safety limitations that we caution against:

  • Secrets: we do not yet have a secrets management solution. So if you have a use case that involves any secrets, the system may read your secrets. All access happens through the OpenAI API – which OpenAI does not train on – but it’s never a good practice to have unencrypted keys go over the open internet. We are exploring some options to integrate with KeyChain / 1Password, etc.

  • Privilege: the system is not running in a container – it has direct access to your system shell. Hence, it is theoretically capable of deleting files or changing configurations on your machine. We’ve never experienced this in testing, but we also do not have explicit safeguards against it. We are working through some ways to address this issue and would love your thoughts on them if you’re interested in discussing it with us.

How to get started

1. Download the app: Download from here: memex.tech

2. Join our Discord: Join other PRP users to explore walkthroughs, get help troubleshooting, and build community: https://discord.gg/DZQDRDuk3m

3. Explore FAQs: We are always available in the Discord, but explore some FAQs here: docs.memex.tech

We look forward to your feedback!

Ever forward,

Memex Team