Extracting Long Conversation States with LoRA

Small-Inevitable6185 · reddit · 2026-07-12

The author is working on a LoRA project for long conversation continuity, aiming not to summarize the entire chat, but to extract a structured "conversation state" from each chunk and merge them into context for subsequent generation.

The current workflow roughly involves:

Real-world engineering data was used, including GitHub Issues, GitHub Discussions, Reddit engineering threads, and long AI dev conversations. Several common clusters were identified: context/memory management, state persistence, reliability, provider compatibility, agent orchestration, long debugging sessions, and architecture discussions.

The author is currently debating whether to invest more in datasets, coverage, and annotation evaluation, or to drop fine-tuning entirely in favor of a stronger base model with prompting.

Original post →

More from coding & agent

coding & agent channel →