EXPERIMENT_010 // AGENT.DREAM.STATE.VISUALISER
Anthropic shipped Dreaming — between-session memory consolidation for Claude agents — in May 2026. This experiment shows what that process might look like as a generative canvas. A scrolling log feed from Diana's real agent fleet drives a background reflection cycle: every 12 seconds, a structured reflection arrives and the particle system shifts. Colour, velocity, and flow encode what the model noticed. The viewer reads that something is consolidating. They cannot name what.
LOADING EXPERIMENT...
Anthropic’s Dreaming feature performs between-session memory consolidation: the agent reviews prior task logs and restructures its stored context so future sessions start from a cleaner, more useful state. The process is invisible. This experiment makes it visible — not as a dashboard or chart, but as an abstract generative canvas that encodes the reflection in colour, particle density, and flow direction. A curated log feed from Diana’s 12-agent fleet cycles through the left panel. Every 12 seconds, the most recent entries are sent to the model for analysis. A structured reflection returns. The canvas shifts. The viewer can read that something is happening. They cannot immediately name what.
HOW IT WORKS
The left panel renders a curated dataset of ~60 realistic log entries drawn from Diana’s actual agent fleet operations: manifest audits, Redis reconnects, worktree lifecycle events, LLM call latencies, SSE stall detections, twin context syncs. The feed advances by one entry every 3–5 seconds, cycling through the dataset. A new entry every few seconds keeps the feed alive without requiring a live fleet connection.
A configurable timer (default 12 seconds, adjustable 8–30s via slider) fires a POST to the reflection API with the last 8 log entries as context. The model returns a structured reflection: dominant signal type (latency, failure, routing, success, or anomaly), intensity, fleet sentiment, whether a memory consolidation update was detected, and a 1–2 sentence observation. The reflection is Zod-validated on the client before it reaches the canvas. A ‘Reflect now’ button triggers an immediate out-of-cycle reflection.
A p5.js particle system maps the reflection to visual parameters: dominantSignal selects the colour palette (amber for latency, red-violet for failure, steel-blue for routing, olive-to-chartreuse for success, near-black for anomaly); intensity drives particle count and velocity; sentiment controls flow turbulence (stable = laminar, degraded = chaotic); tokenEstimate maps to trail length; heuristicUpdate triggers a one-time radial pulse from the canvas centre. Palette transitions interpolate over 2 seconds using p5’s lerpColor — no hard cuts. With prefers-reduced-motion enabled, the particle system stops and a static colour field renders instead.
WHAT THIS PROVES
Anthropic’s Dreaming mechanism consolidates agent memory between sessions — a process with no production-facing visualisation. This experiment demonstrates that a brief log window is sufficient signal for a model to identify dominant operational patterns, assess fleet sentiment, and return a structured reflection that maps cleanly to visual parameters. The model does not need full session transcripts; 8 entries are enough to notice whether the fleet is routing, failing, or recovering.
The architecture also proves that a generative art system can be driven by structured LLM output without any manual parameter tuning. The same DreamReflection schema that governs the API response governs the canvas — the model’s interpretation of the log window is the artistic direction. Every visual shift is grounded in something the model actually noticed.