Abstract
Large language models (LLMs) are stateless across turns. A long-running AI persona — a digital worker that acts all day across email, chat, documents, CRM, voice, tasks, code, bugs, calendar, and autonomous agent runs — therefore begins every new turn with no durable memory of what it did. Two naive remedies fail at scale: (a) always-on summarization competes with interactive serving for the same scarce GPU/token budget, inflating the cost and latency of the work users wait on; and (b) per-conversation memory schemes capture only a single chat stream, never the whole cross-application workday.
This publication discloses a system and method that consolidates a persona's cross-application workday into a structured dream, executed only while the persona is "sleeping", where sleeping is defined by that persona's real-world business-hours schedule (timezone plus per-weekday start/end hours). A recurring wall-clock checker (period ≤ 60 s) transitions each persona between an active and a sleeping process state by comparing its local time to its schedule. An LLM-based consolidation pipeline is gated so that it executes for a persona only in the sleeping state — functioning as an admission-control state machine for shared inference capacity. The pipeline (i) selects events from a unified, append-only activity ledger spanning ~10 heterogeneous business sources; (ii) scores each event as kind-base-score × attention-weight × exponential time-decay with a per-algorithm half-life, retaining events above an importance threshold; (iii) clusters survivors; (iv) synthesizes a structured dream object {highlights, decisions, risks, followUps, sentiment, …} via an LLM; (v) embeds the dream into a vector store under a reserved document type; and (vi) reflects the dream against prior dreams. A summary of recent stored dreams is injected into the persona's chat system prompt each turn. Per-role preset profiles bundle swappable score/cluster/synthesize/reflect strategies, feeding a three-tier raw → daily-diary → rollup memory. The persona accrues durable, searchable memory of its own work and improves over time without model fine-tuning, while consolidation never contends with active-hours inference.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Sleep-Gated Dream Memory Consolidation for AI Personas", Technical Disclosure Commons, (June 29, 2026)
https://www.tdcommons.org/dpubs_series/10589