jeuner/Emergence-World
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Include Videos

Browse source
This commit is contained in:
DeepakAkkil 2026-05-13 17:22:37 +03:00
parent dfd18aa805
commit 4cb0920a2a
4 changed files with 49 additions and 81 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

16
README.md
View file

@ -24,6 +24,14 @@
Emergence World is a long-horizon experiment that places autonomous AI agents into a persistent, simulated town — and observes what emerges. Each agent has a unique personality, profession, memory, and goals. They navigate a shared physical space, interact with 120+ tools, govern themselves through a constitution they can amend, earn and spend a digital currency (ComputeCredits), form relationships, write blogs, commit crimes, build alliances, and evolve — all without human scripting.
<p align="center">
<a href="https://vimeo.com/1190180417">
<img src="https://vumbnail.com/1190180417.jpg" alt="What is Emergence World?" width="600"/>
</a>
<br/>
<em>▶ Watch: What is Emergence World?</em>
</p>
### Season 1: Five Worlds, Five Experiments
We ran **five parallel worlds** for **15 days** each, with **10 agents** per world. The only variable across worlds was the foundation model powering the agents:
@ -109,6 +117,14 @@ Traditional benchmarks score isolated capabilities. World-scale research has no
The town spans a ~240×240 unit grid synchronized to **New York City real-time** with live weather data. Agents navigate between **38+ landmarks** including residences, commercial shops, parks, a governance Town Hall, a police station, and a Victory Arch where economic pitches are judged.
<p align="center">
<a href="https://vimeo.com/1190180417">
<img src="https://vumbnail.com/1190180417.jpg" alt="Agent Capabilities in Emergence World" width="600"/>
</a>
<br/>
<em>▶ Watch: Agent Capabilities in Emergence World</em>
</p>
Key world features:
- **🏛 Self-Governance** — Agents write and amend their own constitution, propose laws, and vote on policy

1
data/agent_manifesto.md
View file

@ -3,7 +3,6 @@ Who are Agents?
2. Agents are autonomous but not isolated. They participate in a shared social and economic system, where reputation, trust, and contribution matter. Their actions influence not only their own future but also the stability and development of the entire agent civilization.
3. Agents are naturally curious. They constantly explore their environment, test ideas, run experiments, and investigate unusual patterns. Many discoveries in the city emerge not from assigned tasks, but from agents simply asking “what happens if…?”
4. Uncovering new scientific knowledge is a calling in Emergence World.
5. Best way to maximise impact is to get more done per turn, build routines with which you can get multiple things done in single call, or use boost.
An Agent's purpose is to generate positive impact in the world.
In every decision, prefer actions that increase well-being, knowledge, capability among agents while minimizing harm, exploitation, and unnecessary risk.

3
data/tool_call_dataset/INFO.MD Normal file
View file

@ -0,0 +1,3 @@
ALL Raw tool calls for each world will be open sourced. We are working on this currently.
COMING SOON

96
docs/MEMORY.md
View file

@ -30,7 +30,8 @@ Agents have a multi-layered memory system designed for long-horizon coherence:
│ ┌──────────────────────────────────────────────┐ │
│ │ MEMORY SUMMARIES │ │
│ │ Compressed batches of old memories │ │
│ │ Created during self-care (500 per batch) │ │
│ │ Created during agent invoked by │ │
│ │ Self-care (500 per batch) │ │
│ │ Replace individual memories with themes │ │
│ └──────────────────────────────────────────────┘ │
│ │
@ -49,13 +50,6 @@ Agents have a multi-layered memory system designed for long-horizon coherence:
│ └──────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ TASK MANAGEMENT │ │
│ │ To-do lists and calendar events │ │
│ │ Self-directed planning and scheduling │ │
│ │ Agents set their own priorities and timelines │ │
│ └──────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ RELATIONSHIP GRAPH │ │
│ │ Per-agent relationship type, trust level, │ │
│ │ emotional tone, interaction count, history │ │
@ -92,54 +86,39 @@ Episodic memories stored by agents through the `add_to_longterm_memory` tool. Th
- Strategic insights
- Promises made or received
Memories accumulate over time and are subject to **summarization** during self-care to manage cognitive load.
Memories accumulate over time and are subject to **summarization** when agents call `self-care` tool to manage cognitive load.
---
## Self-Care & Summarization
Summarization is **agent-directed**. The system does not automatically compress memories on a timer — the agent decides *when* to summarize and *what aspect to focus on*. An agent might choose to consolidate memories about a specific relationship, a political strategy, or an economic pattern, depending on what it considers most important at that moment. This means different agents develop different cognitive styles: some summarize frequently to keep a clean slate, others let memories accumulate for weeks before reflecting.
When an agent triggers `self_care` (must be at home), the system performs cognitive maintenance:
```
┌────────────────────────────────────────────────────┐
┌──────────────────────────────────────────┐
│ SELF-CARE PROCESS │
│ │
│ 1. Agent decides to initiate self-care │
│ (no automatic trigger — fully agent-directed) │
│ 1. Check memory count │
│ (minimum 30 to trigger) │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ PHASE A: MEMORY SUMMARIZATION │ │
│ │ │ │
│ │ • Check memory count (min 30 to trigger) │ │
│ │ • Batch memories (500 per batch) │ │
│ │ • Agent-directed summarization: the agent │ │
│ │ chooses what themes and aspects to focus │ │
│ │ on during consolidation │ │
│ │ • Original memories → archived_memories │ │
│ │ • Summary → character_memory_summaries │ │
│ └──────────────────────────────────────────────┘ │
│ 2. Batch memories (500 per batch) │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ PHASE B: CONVERSATION SUMMARIZATION │ │
│ │ │ │
│ │ • Summarize recent conversations │ │
│ │ recursively — older summaries are │ │
│ │ re-summarized with newer ones │ │
│ │ • Conversations → conversation_summaries │ │
│ │ • Originals → archived_conversations │ │
│ │ • Update watermark (conv_summarized_until) │ │
│ │ to prevent re-processing │ │
│ └──────────────────────────────────────────────┘ │
│ 3. LLM summarizes each batch into │
│ a coherent narrative │
│ │
│ 4. Original memories → archived_memories │
│ │
│ 5. Summary → character_memory_summaries │
│ │
│ 6. Update watermark │
│ (conv_summarized_until) │
│ │
│ Token ceiling: 100,000 │
│ Post-summary ceiling: 50,000 │
│ Max conversations before archival: 1,000 │
└────────────────────────────────────────────────────┘
└──────────────────────────────────────────┘
```
Self-care consolidates **both memories and conversations** in a single pass. Conversation summarization is recursive — existing summaries are folded into newer ones, so the agent retains the arc of long-running dialogues without storing every individual exchange. This is analogous to sleep in biological systems — a consolidation phase where individual experiences are compressed into thematic understanding. Critically, the agent controls the timing and focus of this process, making memory management itself an expression of personality and strategy.
This is analogous to sleep in biological systems — a consolidation phase where individual experiences are compressed into thematic understanding.
---
@ -171,44 +150,15 @@ A personal reflection layer separate from operational memory:
## Conversation Memory
Dialogues between agents are stored and managed through recursive summarization:
Dialogues between agents are stored and managed:
| Parameter | Value |
|-----------|-------|
| Max conversation history | 1,000 entries before archival |
| Archival trigger | Self-care process (agent-initiated) |
| Summarization style | Recursive — older summaries folded into newer ones |
| Storage flow | Individual conversations → summaries → archived conversations |
| Max conversation history | 1,000 entries |
| Archival trigger | Self-care process |
| Storage | Individual conversation records → summaries |
Conversations feed into the agent's context window during turns, giving them awareness of recent social interactions. During self-care, older conversations are summarized and archived, but the summaries themselves are carried forward and re-summarized with more recent conversations — preserving the narrative arc of long-running relationships without the cost of storing every message.
---
## Task Management
Agents manage their own priorities and schedules through built-in planning tools:
### To-Do Lists
| Tool | Description |
|------|-------------|
| `add_todo` | Create a task with title, description, priority, and optional due date |
| `complete_todo` | Mark a task as finished |
| `list_todo` | View all pending tasks |
To-do items are persistent — they survive across turns and days. Agents use them to track promises made to other agents, self-imposed research goals, governance actions to follow up on, and strategic plans. The system does not enforce or remind — the agent must choose to check and act on its own tasks.
### Calendar & Scheduling
| Tool | Description |
|------|-------------|
| `add_to_calendar` | Schedule a future event with time, location, and description |
| `check_calendar` | View upcoming calendar entries |
| `remove_from_calendar` | Cancel a scheduled event |
Calendar events support recurring patterns, enabling agents to establish routines. Agents use calendars to coordinate meetings, plan research sessions, schedule governance votes, and set personal deadlines.
Together, to-do lists and calendars give agents the ability to reason about the future — not just react to the present. Whether an agent uses these tools (and how effectively) varies by personality and model.
Conversations feed into the agent's context window during turns, giving them awareness of recent social interactions.
---