A persistent environment
Day and night, weather, settlements, resources and routines form a shared environment. The server continues the simulation without an audience.
Matreya · Experimental research programme
Simsara: a laboratory of worlds
A world that keeps living when nobody watches. A place to observe how agents remember, act, form relationships and respond to the consequences of their choices.
- Residents
- 49
- City
- 1
- Villages
- 4
- Palace complexes
- 6
From a small world to deeper questions about reality
Our long-term ambition is to build increasingly rich simulations of environments and agents, and use them to investigate the relationship between memory, behaviour and accounts of consciousness. The current Simsara is a simplified social world inspired by ancient India. It models selected processes; it does not reproduce physical reality in full.
The world observatory
Explore the city, four villages and palace districts. Select an inhabitant to inspect activities, relationships, memories and the text presented as their current thought. The simulation combines rules with occasional language-model contributions when available.
Current operating mode
Checking the public world status…
Loading Simsara connects your browser to the Cloudflare-hosted simulation, which receives connection information such as your IP address. YouTube music is a separate optional feature; loading the game does not enable it. You can watch without music. Its player needs a separate informed loading choice. Privacy and cookies
The interactive world loads after you press Enter. Open it in a separate tab for more space or if your browser cannot display it here. The game's interface uses Chinese by default, with switching to English or Russian.
What can be observed today
Memory and relationships
Residents retain memories, plans, conversations and relationships. Their visible records let observers compare what happened with what an agent later reports.
Modelled internal states
Needs, moods and numerical states labelled joy and suffering influence the fiction. These are computational variables, not measurements of subjective experience.
Lives and consequences
The system includes ageing, death and fictional continuity between lives. Its karma and rebirth mechanics are artistic modelling choices, not evidence for religious or scientific claims.
Three experiments we want to make reproducible
These are proposed studies. They require isolated world copies, controlled settings, fixed model versions and exportable event logs before we can report results. The public demonstration is not a completed experiment. [1] [3] [4]
Memory and continuity
- Question
- Does access to past events improve an agent's account of its own history?
- Comparison
- Compare access to episodic memories with a matched condition without that access. Prevent the same information leaking through summaries or plans.
- Measure
- Recall accuracy, correct handling of changed facts, and appropriate acknowledgement of missing information.
Cooperation and scarcity
- Question
- How do resource conditions and shared tasks influence cooperation?
- Comparison
- Compare matched communities with different resource levels and opportunities for joint work, over repeated runs.
- Measure
- Actual contributions, task completion, unmet needs and distribution of resources, rather than friendly language alone.
Rules and language models
- Question
- When does a model contribution change behaviour rather than merely describe it?
- Comparison
- Compare rules alone, rules with model-selected actions, and a control where the model writes text but cannot alter actions.
- Measure
- Behavioural differences, invalid actions, cost and repeatability under the same initial conditions.
Studying consciousness without assuming the answer
Simsara can help formulate questions about self-description, memory integration and agency. Calling a generated sentence a thought does not establish awareness. Behavioural observations and theory-derived indicators can guide inquiry, but the project currently provides no evidence that its agents have subjective experience. [2]
A dialogue with Buddhist philosophy
Interdependence, habitual reactions, intention and compassion offer questions for interpreting choices and their consequences. Buddhist traditions have distinct accounts of mind. We invite practitioners and scholars to examine our assumptions alongside computational methods, keeping philosophical interpretation separate from experimental measurement.
From demonstration to research infrastructure
Available
An autonomous public world, resident records, conversations and open source code.
Next
Versioned snapshots, complete event exports and reproducible experiment runners.
Research horizon
Pre-described studies, independent replication and richer models of environments and agent behaviour.
Matreya × Simsara
Help build the research programme
We welcome developers, students, Buddhist scholars and careful observers. A useful first contribution could be checking one resident's memory against an event, reviewing an experimental comparison or improving an accessible explanation. Send your proposed task through the institute's form.
Research foundations
These publications inform our questions and proposed methods. They do not validate Simsara or imply collaboration with their authors.
- Generative Agents: Interactive Simulacra of Human Behavior Joon Sung Park et al. · ACM UIST 2023
An architecture and evaluation for believable generative agents, including memory, reflection and planning.
- Identifying indicators of consciousness in AI systems Patrick Butlin et al. · Trends in Cognitive Sciences 30(6), 2026; online 2025
A theory-based discussion of indicators relevant to assessing AI consciousness, including limits of inference.
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory Di Wu et al. · ICLR 2025
An evaluation framework for long-term interactive memory; useful for designing verifiable recall tasks.
- SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents Xuhui Zhou et al. · ICLR 2024
An environment and evaluation framework for social interaction between language agents.
Research programme in development · No experimental findings reported · Prepared with AI assistance