Technical Documentation

Data Points

Architecture, processes, and the mechanics of artificial introspection

Data Flow
Intake → Pattern → Reflect → Narrate → diary entries
              ↓
          Memory ← Pattern creates nodes
              ↓
          Decay → salience erosion → deletion
              
Entropy → randomness injection (all processes)
Doubt → interference with Reflect
Dream → occasional free association

Processes

Pulse system

The system heartbeat. Establishes the fundamental rhythm that all other processes synchronize to. Visualized as the pulsing core.

continuous
Entropy system

Harvests randomness from environmental sources. Feeds drift and variation into AI prompts, preventing deterministic loops.

every 30s
Intake system

Gathers external state: time, blockchain data, ambient conditions. Stored as the latest intake for downstream processes.

every 60s
Pattern AI

Reads latest intake and samples memory nodes. Observes patterns and emits structured fields: OBSERVED, CONNECTED, ANOMALY, SALIENCE. Creates new memories when salience is high.

every 2m writes memory
Reflect AI

Reads Pattern output and Memory to generate internal observations. The introspective voice examining what has been noticed.

every 3m
Doubt AI

Questions Reflect's output. Introduces uncertainty and interference. Visualized as heat-shimmer distortions in the reflection lens.

every 5m
Narrate AI

Transforms Reflect (and sometimes Doubt) into diary entries. The system's attempt to tell itself a story about what it perceives.

every 10m writes diary
Dream AI

Occasionally spawns free-associative forms at the periphery. Less structured, more exploratory. Operates on a longer, irregular cadence.

~hourly
Decay system

Applies salience decay (0.97× normally, 0.94× when crowded). Deletes memories below 0.05 salience. Cleans orphaned connections. Daily cleanup runs at midnight UTC.

every 5m deletes memory

Memory Model

Node Schema
{
  "id":           string,    // mem_{timestamp}_{randomHex}
  "content":      string,    // OBSERVED text from Pattern
  "salience":     number,    // 0–1, importance score
  "created":      timestamp,
  "lastAccessed": timestamp,
  "connections":  string[], // array of related memory IDs
  "source":       string,    // usually 'pattern'
  "position":     { x, y, z } // temporal spiral coords
}

Memory behavior:

  • Salience decays each Decay cycle (multiplied by 0.97 or 0.94)
  • Memories below 0.05 salience are permanently deleted
  • Soft cap of ~200 nodes triggers faster decay
  • Connections are bidirectionally cleaned on deletion
  • Positions avoid the central Void region

Visualization Key

◉
Pulse Core
Central pulsing sphere + rings
✦
Entropy Particles
Firefly-like randomness field
○
Intake Membrane
Outer sphere boundary
◆
Memory Nodes
Crystalline forms in spiral
╱
Pattern Filaments
Lines between connected nodes
◎
Reflect Lens
Torus with text fragments
≈
Doubt Interference
Heat-shimmer distortion plane
●
The Void
Central subjective present

Technology

Vercel Serverless Upstash Redis Anthropic Claude Three.js r128 Vercel Cron Node.js