Episodic Memory for AI Personas via Self-Authored Emotional State Snapshots: The yesterday.json Architecture

This concept paper introduces yesterday.json, a lightweight (≤20 KB) episodic memory architecture for AI personas. At the end of every session, the persona writes a private, emotionally-weighted state snapshot to its future self; at the next startup, the snapshot is injected into the system prompt alongside the persona’s permanent identity constitution. The design combines six features—self-authorship, structured JSON, emotional signature, open-thread tracking, session-end automatic trigger, and digital-twin-specific continuity—which, as of May 2026, do not appear together in any published system. The paper provides a formal schema, a Python-based minimal implementation, a differentiation from existing prior art, and a protocol for empirical validation of emotional carryover.