🚨Anthropic leak details Claude Code memory index and retrieval system

Hey ,

Your daily briefing is ready. You can finally take a break from the AI firehose.

Our algos spent the night splitting signal from noise and pulled the top news, models, papers, and repos.

Here’s the must-read:

Together with:

Summary

Read time: 4 min 24 sec

Top Repo

▸ Anthropic’s Claude Code gets leaked revealing 44 feature flags and projects

WorkOS

▸ Connect your app to Slack, GitHub, and Salesforce using one API

Top News

▸ Google releases Veo 3.1 Lite with lower-cost video generation

Signals

▸ Google publishes a method to design better and cheaper AI benchmarks

▸ Hugging Face presents TRL v1, a post-training library for open modelsn

▸ Liquid AI unveils an open-weight edge model optimized for tool use

▸ Salesforce updates Slackbot agent with 30 new capabilities in Slack

▸ Perplexity creates institute to study security, privacy, and safety of AI systems

Today's Author

Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.

Top Repo

Anthropic's Claude Code leak shows a memory system that stores pointers, rewrites itself, and avoids context overload

101,528 Stars

Anthropic accidentally exposed Claude Code’s internals, including ~1,900 files and a full agent runtime.

At first glance, it looks like another AI coding tool. But the key detail sits in how it handles memory: it does not store knowledge directly, it stores references.

Most AI systems try to remember everything and hit context limits fast.

Claude Code avoids that by loading a tiny index (MEMORY.md) into every prompt, where each line points to external files. When you ask something, the system fetches only the relevant file instead of dumping everything into the model.

Then comes the unusual part: memory does not grow, it gets rewritten. A background process merges duplicates, removes contradictions, and deletes outdated facts. The system treats stale memory as incorrect and fixes it continuously.

The leaked architecture also shows a full agent harness, replicated in claw-code GitHub repository, with runtime, tools, and CLI layers.

Key Insights

Pointer-based memory keeps prompts small and avoids unnecessary token usage
Three-layer retrieval loads index always, files on demand, transcripts via search
Continuous rewriting maintains consistency instead of appending conflicting entries
Modular harness supports tools, plugins, commands, and session-level orchestration

STUDY THE CODE

Presented by WorkOS

Unlock access to the data your product needs

Most tools are still locked to their own database, blind to everything users already have in Slack, GitHub, Salesforce, Google Drive, and dozens of other apps. That's the ceiling on what you can build.

WorkOS Pipes removes it. One API call connects your product to the apps your users live in.

Pull context from their tools, cross-reference data across silos, power AI agents that act across services. All with fresh, managed credentials you never have to think about.

GET STARTED

partner with us

Top Paper

Stanford study reveals multimodal models maintain 70-80% accuracy even after complete removal of image inputs

5,483 Likes

Stanford researchers release MIRAGE, a method to test how multimodal models use images. They take GPT-5.1, Gemini-3-Pro, Claude Opus 4.5, and Gemini-2.5-Pro and remove every image from six benchmarks.

The surprising part: models still keep 70-80% accuracy and generate detailed visual reasoning as if images exist.

The problem starts with how benchmarks work. Many questions contain enough text clues to infer answers without vision. Models learn these patterns and rely on them during inference instead of processing pixels.

The paper calls this behavior “mirage reasoning.” The model assumes an image exists and constructs an answer using text correlations. A 3B text-only model, trained on 696,000 chest X-ray QA pairs without images, ranks first on the test set.

Key Results

Models maintain high accuracy without visual input across multiple benchmarks
B-Clean removes 74-77% of samples that do not require actual vision
Text-only model outperforms larger multimodal systems on the same dataset
Outputs include structured reasoning that mimics visual analysis

READ THE PAPER

Presented by Unblocked

Cut rework and token usage with better context

Stop babysitting your coding agents. Unblocked gives them the organizational knowledge to generate mergeable code without the back and forth.

It pulls context from across your engineering stack, resolves conflicts, and cuts the rework cycle by delivering only what agents need for the task at hand.

UNBLOCK YOUR AGENT

partner with us

Top News

Google introduces Veo 3.1 Lite with lower cost video generation and same speed as Fast model

2,492 Likes

Google DeepMind releases Veo 3.1 Lite, a lower-cost video generation model in the Gemini API. The model costs less than 50% of Veo 3.1 Fast while maintaining the same generation speed. Google also schedules a price reduction for Veo 3.1 Fast on April 7.

Model scope

Veo 3.1 Lite supports standard video generation workflows and output configurations.

Handles both text-to-video and image-to-video generation tasks within a single model
Supports landscape 16:9 and portrait 9:16 output formats for different use cases
Generates video at 720p and 1080p resolutions based on selected configuration
Provides configurable clip durations of 4 seconds, 6 seconds, or 8 seconds

Usage access

You can access and run the model through existing Google developer tools.

Available via paid tier access in Gemini API for programmatic video generation
Integrated into Google AI Studio for testing prompts and iterating outputs
Accepts text prompts or input images to generate short video clips
Requires selecting resolution, format, and duration parameters per request

TRY VEO LITE

Signals

Google builds a framework to choose between more data points or more raters for reliable AI evaluation 2,428 Likes

Hugging Face announces TRL v1, a post-training library with over 75 methods including SFT, DPO, and R 2,401 Likes

Liquid AI releases LFM2.5-350M, an open-weight 350M model optimized for fast edge inference 1,837 Likes

Salesforce upgrades Slackbot to execute workflows across apps, desktop, and enterprise data in chat 1,423 Likes

Perplexity creates SII to study defenses and risks in agents, models, and real-world AI deployments 1,023 Likes

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 250,000+ AI developers?

WORK WITH US

How was today’s email?

Awesome Decent Not Great

unsubscribe_me(): return True

🚨 Anthropic Claude Code leak reveals pointer-based memory system