Detailed Analysis
A Reddit user's question about Claude's Projects feature highlights a persistent point of confusion among users trying to manage multi-session workflows with Anthropic's AI assistant. The poster describes setting up a Claude Project—Anthropic's mechanism for grouping conversations around shared documents, instructions, and context—only to find that new chats within that same project behave as if they have no memory of previous conversations. Despite having reference materials stored in the project's knowledge base, each new chat thread starts essentially from scratch, unable to recall decisions, context, or progress from earlier sessions unless the user manually locates and reopens that exact prior chat.
This limitation stems from a fundamental aspect of how Claude Projects actually function versus how users often expect them to work. Projects provide a shared knowledge base (uploaded files, custom instructions, and a system prompt) that persists across all chats within that project, but they do not automatically carry conversational history or "memory" from one chat thread to another. Each conversation is still a discrete, stateless session as far as the model's active context window is concerned. The project's static documents are visible to every new chat, but anything discussed, decided, or generated in Chat A is invisible to Chat B unless the user explicitly copies that information into the project's knowledge files or pastes relevant excerpts into the new conversation. This is a common source of frustration because the branding and UI of "Projects" implies a persistent workspace with continuity, when in practice it behaves more like a shared folder of static reference material attached to otherwise isolated chats.
This distinction matters because it shapes how professionals and power users can realistically deploy Claude for long-running, multi-session work like research projects, coding efforts, or ongoing writing collaborations. Users accustomed to human collaborators or to tools with persistent memory expect an assistant to "remember" what happened last week, but Claude's architecture—like most current large language models—relies on context windows that reset with each new conversation unless explicitly managed. The practical workaround, which experienced users often adopt, involves maintaining a running summary document (a "state of the project" file) that gets manually updated and uploaded to the project's knowledge base after each session, effectively simulating memory by offloading it to persistent storage rather than relying on the model itself to retain it.
This gap also reflects a broader industry-wide challenge: the tension between stateless, context-window-based AI systems and the growing user demand for persistent, evolving memory across interactions. Anthropic, OpenAI, and others have begun experimenting with features like cross-chat memory (as seen in ChatGPT's memory feature) to address exactly this pain point, but Claude's Projects, as of this writing, remain closer to a curated knowledge repository than a true continuous-memory system. As AI assistants are increasingly used for sustained, complex work rather than one-off queries, the ability to preserve and surface relevant context automatically—without requiring users to manually reconstruct it—will likely become a key differentiator among competing AI platforms, and a natural area for Anthropic to expand upon in future Claude updates.
Read original article →