Claude's Source Code Got Leaked Across The Whole Internet
Anthropic's Claude source code was leaked across GitHub, revealing details about how the AI system handles memory and context storage. The leak showed that Claude uses an indexing system that stores location references rather than raw data, and searches through transcripts rather than fully loading them into context.
Summary
The video discusses a significant leak of Anthropic's Claude AI source code that spread widely across GitHub. The speaker focuses on the technical insights gained from examining the leaked code rather than the leak itself. The most notable discovery relates to Claude's memory and context management system. According to the leaked code, Claude employs an indexing system that stores location references rather than the actual data itself. When users interact with Claude, the system saves information about the conversation and creates references to this data in a memory file. Rather than loading complete transcripts back into the context window, the system uses a 'grep' function to search for specific identifiers within the stored transcripts. The speaker explains that when Claude needs to recall information, the memory file tells the context system where to search for relevant information rather than loading everything directly. The speaker acknowledges they may be oversimplifying the technical details but presents this as their understanding of how Claude's memory architecture functions based on the leaked code.
Key Insights
- Anthropic's Claude source code was leaked and spread widely across GitHub
- Claude's indexing system stores location references rather than storing the actual data itself
- Raw transcripts are never fully loaded back into context but are instead searched using grep functions for specific identifiers
- Claude saves conversation information and creates references to this data in a memory file during interactions
- The memory file directs the context system to specific search locations rather than loading complete conversation histories
Topics
Transcript
spread across GitHub like crazy, Anthropics' clod code was actually leaked onto the internet. I believe this is the original source of the discovery here, but I think the most interesting thing to come out of this leak is actually what we learned from looking at the code and what was actually in the leak. The index does not store data, it stores locations. While raw transcripts are never fully read back into the context, but merely grepped for specific identifiers. Now, grep is a word that essentially means that it searches. So as you're talking to clod code, it's saving a bunch of information about what you're talking to it about, and then it's saving a reference to all…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Matt Wolfe
The ONLY AI Benchmark You Need!
A developer created "Buccy Bench," a humorous yet functional AI benchmark that tasks different language models with drawing Gary Busey as SVG code rather than images. The benchmark tracks model evolution over time while measuring performance metrics like cost, tokens, and execution speed.
GLM-5.2 - The Open Model That's As Good As Opus!
A comprehensive review of GLM-5.2, an open-weight Chinese AI model with a 1 million token context window, demonstrating its capabilities for coding, document analysis, and agentic workflows at significantly lower costs than frontier models like Claude Opus and GPT-4.5. The speaker tests various use cases including website building, Chrome extensions, game development, and data organization, concluding it's valuable for long, code-heavy, token-expensive tasks despite not universally outperforming closed-source alternatives.
Don't Fall For This AI Trap
The speaker emphasizes that power users distinguish themselves by knowing what NOT to automate with AI, rather than automating everything. They argue that AI works best for clear, straightforward tasks but struggles with nuanced, artistic work requiring consistency—using their failed YouTube thumbnail automation as an example.
AI News: Fable Banned, New Open-Source Leader, Midjourney Shocker
This AI news roundup covers the US government forcing Anthropic to shut down its Fable 5 and Mythos 5 models due to a security vulnerability jailbreak, the release of a competitive open-source model GLM 5.2 by ZAI, and MidJourney's surprising pivot into medical imaging technology with a new ultrasound-based body scanner.
AI News: Claude's Massive Leap & Siri Gets Good!?
This AI news roundup covers the release of Claude Fable 5 (a Mythos-tier model from Anthropic) and its controversial safety restrictions, Apple's WWDC AI announcements including a major Siri overhaul, and updates from Google including NotebookLM upgrades and real-time translation. Additional rapid-fire items include OpenAI and SpaceX IPO filings, ChatGPT email sending, and a teased Midjourney hardware device.