Three Storage Tiers and Their Contents

STM · recent turnsqueue of pages · FIFOoldest page spills · match scoreMTM · topic folderspages by topic · heat · evictheat > τ · promoteLPM · personauser · agentretrieveall three tiers → promptgenerateLLM → answer → new page
STM · recent turnssnapshot · week 1, right after p3 page = {Q, R, T} · pages on one thread form a chain chain p1T · week 1, Mon 19:02 Q went to the wetland park, scenery beautiful R sounds like a relaxing experience p2T · week 1, Mon 19:04 Q many little squirrels in the trees, very cute R squirrels signal a healthy environment p3T · week 1, Mon 19:07 Q ran two laps, will keep working out to get slimmer R exercise is good, I'll keep you motivated chain meta · running summary, written by an LLM "user visited a wetland park, saw squirrels, ran two laps, wants to get slimmer" L MTM · topic folderssnapshot · week 4, after p1–p3 left the queue folder (segment) = pages on one topic + summary + keywords wetland park outing heat · hot summary L "a park visit with squirrels and a run; the user wants to get in shape" keywords park squirrels run slimmer pages p1park p2squirrels p3run work deadlines · 2 pages · warm old laptop issue · 1 page · cold LPM · who the user and the agent aresnapshot · week 4, after the folder was promoted fixed fields stay · learned fields grow (FIFO queues of 100) User Persona Profile static name · gender · birth year User KB dynamic facts learned from the chat · went to the wetland park (week 1) · runs laps in the park User Traits dynamic 90 tags in 3 groups · goal: get slimmer · likes outdoor exercise needs & personality · AI alignment · content interests Agent Persona Profile static role: a friendly health-minded assistant Traits dynamic what it did for this user · promised to keep the user motivated
13 / 25