Current chat systems built on large language models (LLMs) treat the conversation history as an immutable sequence of turns that defines the model's context. In real interactions, user intent is dynamic—it evolves through corrections, refinements, and shifting constraints. This mismatch leads to context pollution: outdated or irrelevant information persists and continues to affect later responses.
We introduce mutable transcripts, an interaction paradigm that lets users revise previous turns via natural‑language edit requests instead of merely appending new ones. The transcript thus becomes an editable representation of conversational state rather than a passive log.
A working prototype embeds transcript‑level revision into a standard chat interface, allowing deletions, replacements, or insertions at any point in the dialogue. Feasibility was evaluated through a controlled user study (n=17) and an analysis of representative interaction transcripts. Participants significantly preferred mutable transcripts over the conventional approach on clarity, confidence, and ease of use, and reported a lower intent to restart conversations. Transcript analysis showed reduced conversation length and elimination of obsolete retained context.
The source code and prototype are publicly available at https://github.com/QxLabIreland/ReChat.
Review