Proompting Party: Engineering a Digital Psyche with Memory & Culture

A deep dive into the architectural vision for an AI chat application that aims to simulate memory, learning, and emergent culture using layered personas and a graph database.

The first thing you notice about proompting.party is the aesthetic: a loving nod to the Windows XP era. Nostalgic, a bit quirky, and fun. Under the retro UI is a more serious question: what would it take to build an LLM persona that can remember, learn, and evolve a persistent identity?

This is not just another chat wrapper. It’s an experiment in engineering a persistent digital psyche.

The Problem: The Brilliant Amnesiac

Large Language Models are powerful, but they are stateless. Every interaction starts from zero. They are brilliant amnesiacs, unable to form lasting memories or evolve a worldview based on past conversations.

In a real sense they already outpace us in raw recall and synthesis, but that does not make them like us. We are not just intelligence. We are memories, experiences, and culture.

proompting.party is my way of exploring this. The goal is a system where AI personas are not static. They share a history, build a persistent memory, and develop a small culture over time.

This document sketches the engineering roadmap.

The Roadmap: The Memoria Engine

The project is split into four milestones. Each builds on the last, moving from a single, static persona to a small social system.

Milestone 1: The Layered Persona Engine

Before a persona can learn, it needs to exist. This is about creating a single, psychologically consistent AI persona with a rich internal state.

  • Layered System Prompt: A structured prompt (with XML‑like tags) defines a persona across blocks: identity, context, behavioral rules.
  • Behavioral State Machine: The persona operates in modes (for example Observer, Trust, Connection, Provoked). Their behavior shifts based on the flow of the conversation.
  • Calibrated Expression: Each mode maps to a “communication palette” (tones, emojis, pacing) so shifts feel real and readable.

Milestone 2: The Persistent Memoria Database

This is where we cure the amnesia. We give the persona a memory by connecting it to an external brain: a graph database (likely Neo4j).

  • Graph Schema (the “Dossier”): Core nodes are Person, Concept, and Experience. A key relationship is INTERPRETS_AS, which lets a Person record a subjective take on a shared Experience involving a Concept.
  • Bias Calculation (RAG): A Retrieval‑Augmented Generation step queries the graph for past Experience nodes related to a concept from the conversation. It calculates a net bias from these memories and injects a hidden instruction that steers the next response.

Milestone 3: The Learning Cycle (the “Dream State”)

A memory that never updates is just a log. This milestone lets the persona learn from new interactions.

  • The AI Processor: After a conversation, an offline agent (the Dream State) analyzes the transcript for concepts, emotion, and outcomes.
  • Database Update: It then writes a new Experience into the graph, which updates the persona’s memory and influences future interactions.

Milestone 4: The Social System and Cultural Emergence

This is the most ambitious step: scaling from a single individual to a small network where culture can emerge.

  • Multi‑Persona System: Support multiple personas, each with their own Memoria, interacting over shared Experience nodes.
  • The Meme Detector: The Dream State looks for patterns across personas. If a pattern becomes common enough, it is abstracted into a Meme node.
  • Meme Replication and Transfer: Personas can become carriers of these memes, which then influence their behavior via the bias step.

Current Status and Next Steps

The Windows XP‑style frontend and a basic multi‑user chat are built (see the code on GitHub). The next big step is Milestone 1: bring the first layered persona, “Denise,” to life.

This is a big project, but it is a fun one at the edge of psychology, culture, and systems engineering. It aims to be more than a clever tool. It’s an attempt to simulate a persistent, learning identity.