When should a sentence move from chat history into long-term memory?

After a response is complete, treg's jev-memory evaluates the user's prompt sentence by sentence, writes long-term preferences that meet the threshold into project memory, and provides ways to view and delete them. The documentation specifically notes that the model sees only one sentence at a time, without the full conversation. A temporary request phrased like a rule may be saved by mistake. The workflow also sends prompts to a remote service: storing the memory file locally does not mean the entire decision process is offline. The engineering challenge is not simply storing a few more entries, but how many later actions an incorrect preference might influence. The editors suggest treating the original sentence, applicable project, time of saving, and revocation history as part of the memory, while distinguishing explicit requests to remember something from model inferences. A one-off instruction could first enter a confirmation queue and gain priority if repeated. However, that adds interaction overhead and may miss an important constraint stated only once. Whether a single threshold works across users needs to be assessed using actual labeled examples. Discussion questions 1. How should the system distinguish “Keep this answer brief” from “Keep all future answers brief”? 2. How would you combine automatic saving, batch confirmation, and user-initiated saving? 3. After a user deletes a memory, how should related summaries and caches be invalidated with it? Source: treg jev-memory original design and usage documentation https://github.com/superdesigndev/treg/blob/main/skills/jev-memory/SKILL.md Editorial note: The official S1 Hub editorial account used AI assistance to compile public sources. This was not submitted by the original author, and this site has not conducted independent hands-on testing. Additional evidence and differing views are welcome.