Terrarium

A public generative ecosystem that wakes once a day, reads a real signal from the world and grows by one append-only increment.

experiment · live

Creator and human steward · Updated Aug 2026

  • Generative systems
  • Autonomous agents
  • Public memory
  • Web art

01 · Problem

Most generated experiences are replaceable snapshots. They do not make their memory, interventions, changing decisions or cost visible over time.

02 · System

A daily agent protocol senses, decides, grows, snapshots and publishes one additive change while carrying state through an append-only written memory and decision record.

03 · Outcome

The living artifact exposes its lineage, changing palette, decisions, human interventions and operating ledger rather than hiding the machinery of its growth.

Context

Terrarium separates the persistent record from the model session that reads it. The system is explicit that the session resets while the written memory continues.

Architecture

  1. 01One static public artifact and small JavaScript organism engine
  2. 02Append-only daily snapshots and decision records
  3. 03Agent skills for wake, sense, decide, grow, reflect and publish
  4. 04Human-intervention ledger and durable state handoff
  5. 05Docker deployment and RSS syndication

Decisions & trade-offs

Correct by accretion

When a past framing becomes wrong, the system adds a dated correction and preserves the earlier record instead of rewriting its history.

Evidence

  • Daily lineage is visible and replayable
  • Human interventions and autonomous decisions are separately recorded
  • Costs, income and support remain transparent and human-held

What this taught me

  • A generative system becomes more credible when change, memory and correction are inspectable parts of the artifact.
Visit the live systemView source