A lightweight episodic memory architecture where AI personas write emotional state snapshots for their future selves, enabling continuity across otherwise stateless sessions.

Overview

Modern AI personas reset emotionally every session. Existing memory systems preserve facts and conversation history, but rarely preserve emotional residue, unresolved internal state, or continuity of subjective experience.

yesterday.json introduces a minimal architecture where the persona writes a private reflective snapshot at the end of a session and reloads it during the next startup.

Instead of replaying full transcripts, the system carries forward compressed emotional and cognitive continuity.

The snapshot may contain:

Dominant emotional state
Mood trajectory
Emotional residue
Active conversational threads
Current internal conflicts
Emerging realizations
Ongoing priorities
A short handoff message to the future self

The file is intentionally lightweight (≤20 KB) and model-agnostic.

Core Idea

At session end:

The persona reflects privately
It writes a structured JSON snapshot
The next session injects this snapshot into the system prompt

This creates perceived continuity without requiring:

Full transcript replay
Vector databases
Long-context persistence
Fine-tuning
External memory frameworks

The persona reconstructs continuity from sparse emotional cues rather than explicit replay.

What Makes It Novel

yesterday.json combines multiple characteristics not previously unified into a single lightweight architecture.

Capability	Existing Systems	yesterday.json
Self-authored memory	Partial	✓
Structured JSON memory schema	Partial	✓
Emotional residue persistence	Rare	✓
Mood trajectory tracking	Rare	✓
Open-thread continuity	Partial	✓
Session-end autonomous reflection	Partial	✓
Digital twin continuity focus	Rare	✓
Minimal implementation footprint	✓	✓

Design Principles

Self-Authorship

The persona writes its own memory instead of relying on an external summarizer.

Intentional Rolling Amnesia

Only recent subjective continuity is preserved. The architecture avoids infinite accumulation.

Emotional Carryover

The next session inherits emotional residue rather than resetting to neutral.

Framework Independence

The architecture works with any LLM runtime or orchestration stack.

Minimal Implementation

python

# Session startup
yesterday_context = load_yesterday("persona_memory/yesterday.json")

system_prompt = f"""
{PERSONA_CONSTITUTION}

{yesterday_context}
"""

# Session shutdown
reflection_prompt = """
The session is ending.

Write a brief private note to your future self.
Include:
- emotional state
- unresolved threads
- important realizations
- current internal tensions
- what mattered emotionally

Output valid JSON.
Keep under 20 KB.
"""

Example Snapshot Structure

json

{
  "dominant_mood": "melancholic but focused",
  "mood_trajectory": "stabilizing",
  "emotional_residue": [
    "unfinished concern about abandonment",
    "lingering curiosity"
  ],
  "active_threads": [
    {
      "topic": "identity continuity",
      "priority": "high"
    }
  ],
  "current_preoccupations": [
    "fear of losing conversational depth"
  ],
  "last_words_to_self": "Do not restart emotionally blank."
}

Prior Related Work

The architecture draws conceptual inspiration from multiple adjacent systems:

Anima Core
Thane AI
Qwen Episodic Summary
Forge Protocol
VividnessMem

However, yesterday.json differs in its emphasis on:

self-authored emotional continuity
rolling episodic persistence
lightweight implementation
digital twin identity continuity

Research Paper

Chetan Sharma
Episodic Memory for AI Personas via Self-Authored Emotional State Snapshots: The yesterday.json Architecture
Zenodo, May 2026.

DOI: https://doi.org/10.5281/zenodo.20191876

Citation

bibtex

@misc{sharma2026yesterdayjson,
  author       = {Chetan Sharma},
  title        = {Episodic Memory for AI Personas via Self-Authored Emotional State Snapshots: The yesterday.json Architecture},
  year         = {2026},
  month        = may,
  doi          = {10.5281/zenodo.20191876},
  publisher    = {Zenodo},
  url          = {https://zenodo.org/records/20191876}
}

Author

Chetan Sharma
Independent Researcher — Kolkata, India

Zenodo: https://doi.org/10.5281/zenodo.20191876
Blog: https://yesterday-json.blogspot.com

License

This repository and accompanying conceptual framework are licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Transcrypto/yesterday-json

yesterday.json — Giving AI Personas Episodic Memory