Claude's Source Code Got Leaked Across The Whole Internet
In this solo commentary, Matt Wolfe explains the recent leak of Anthropic's Claude source code and shares his interpretation of the model's memory and indexing system. He asserts that the index does not hold raw transcript data but rather pointers to locations where relevant information resides. When a conversation occurs, Claude saves contextual details, creates references to those details in a memory file, and later loads that memory file into the active context. The system then uses a grep‑like search to locate specific identifiers within the stored transcripts rather than re‑reading entire texts. Wolfe acknowledges that his description may oversimplify the actual implementation but presents this as the primary insight gleaned from the leaked code.
