Detailed Analysis
A Reddit user working extensively with Claude for deep research has identified and documented a structural failure mode in Claude Projects — Anthropic's organizational feature designed to help users manage complex, ongoing AI-assisted work. The core finding is that individual chats within the same Project do not share conversational context with one another, despite operating under a shared Project interface, shared name, and shared Knowledge folder. The user discovered this when opening a second chat (Chat B) within a Project and receiving a suggestion from Claude that had already been explicitly addressed and resolved in the first chat (Chat A). The chats are, in the user's framing, "context-blind to each other" despite their apparent structural relationship.
The more consequential element of this critique is not the isolation itself, but the behavioral trap it creates for users. When users discover that chats within a Project do not share context, the intuitive workaround is to consolidate everything into a single, continuous conversation — which is precisely the behavior that accelerates AI memory drift, the problem Projects ostensibly exists to solve. The UX design of the Projects feature — with its shared folder, shared workspace, and shared identity — implies a degree of contextual awareness between chats that does not actually exist. This gap between implied and actual architecture means users encounter the failure mode only after investing significant time, and their natural corrective response compounds the original problem rather than resolving it.
The user's proposed fix is deliberately scoped and technically conservative. Rather than requesting full cross-chat conversation history merging — which raises both technical and privacy concerns — the proposal calls for a lightweight shared project-level context layer that all chats within a Project can read. This would function as a persistent, project-wide semantic summary or topic registry, distinct from full conversation logs, enabling Claude to recognize when a new chat is revisiting territory already covered elsewhere in the same Project. The Knowledge folder already demonstrates that Anthropic's architecture supports shared document-level context within Projects; the argument is that extending this principle to a dynamic, AI-generated context summary layer is the logical and tractable next step.
This critique sits at the intersection of two well-documented challenges in contemporary large language model deployment: context window limitations and the design of human-AI collaboration workflows. Memory drift — the degradation of coherence and accuracy in long conversations as earlier context recedes or gets compressed — is a known limitation of transformer-based architectures and a persistent frustration for users conducting multi-session research. Various mitigation strategies have emerged, including retrieval-augmented generation (RAG), external memory systems, and structured summarization, but these are largely invisible to end users of consumer-facing products like Claude.ai. The Projects feature represents Anthropic's attempt to address this at the product layer rather than purely the model layer, which makes the architectural gap identified here particularly significant.
Broadly, the failure mode described reflects a wider tension in AI product design between technical architecture and user mental models. Features that suggest organizational coherence — shared workspaces, project containers, named contexts — create strong expectations of functional coherence that the underlying infrastructure may not yet support. As AI assistants are increasingly positioned as long-term research and knowledge-work partners rather than single-session tools, the demand for persistent, cross-session contextual awareness will intensify. Anthropic faces competitive pressure on this dimension from systems exploring various forms of persistent memory, and the community feedback captured in this post illustrates that users are actively probing the limits of current architectures and articulating specific, technically grounded solutions.
Read original article →