Current AI chat interfaces force conversations into a single, linear thread. But human thinking isn’t linear — we explore tangents, need to clarify specific points without polluting the main thread, and rely on context that we forget to state upfront. The result is long, cluttered conversations where useful content gets buried, the AI loses focus on what matters, and the user has no real control over what the model retains.
These four features, taken together, solve that problem as a cohesive system.
1. Branching Threads When the AI gives a response, the user should be able to open a branch from any specific message to explore a sub-topic — without that tangent appearing in the main conversation. Think of it like a comment thread inside a thread. Once the branch is resolved, the main conversation continues cleanly. This keeps the primary flow readable while still allowing depth.
2. Prompt-Level Context Metadata When the AI detects missing information in a prompt, instead of asking a clarifying question inside the chat (which clutters the timeline), it should open a small popup outside the main thread. The user answers there, and that answer gets attached as invisible metadata to that specific prompt — not as a visible message. Later, the user can right-click that prompt and select “View context” to inspect what metadata is attached. This keeps the conversation clean while making the AI’s reasoning transparent and auditable. This is intentionally different from persistent memory settings — it’s per-prompt, ephemeral, and tied to a specific moment of ambiguity.
3. Right Sidebar with Prompt Navigation A collapsible right sidebar listing all prompts in the conversation, allowing the user to jump directly to any point. This pairs naturally with threads — branches would appear indented under the parent prompt. No more endless scrolling to find something said 40 messages ago.
4. Selective Memory Visibility Each entry in the sidebar that the AI is actively using as context would be highlighted (e.g. in blue). The user could uncheck any entry to exclude it from the context window on the next request — without deleting it from the visible chat history. The message stays on screen, but the AI simply doesn’t receive it in the next payload. Branches could be excluded independently from the main thread. Context metadata (from feature #2) would show a small indicator on the sidebar entry, but wouldn’t be toggled the same way — since it’s already scoped to a single prompt.
Each feature solves one layer of the same underlying problem: conversations are linear but thinking isn’t.
Threads solve depth without pollution. Context metadata solves intent without verbosity. The sidebar solves navigation. Selective memory solves control over what the AI actually knows at any given moment.
Separately, each is a nice quality-of-life improvement. Together, they turn the chat interface into something closer to a structured thinking environment — powerful for researchers, students, and power users, while remaining opt-in and non-intrusive for casual use.