Reverie

简体中文 | English

Reverie is an agent runtime for self-evolution: when nobody asks, it keeps exploring and growing; when asked, it answers from its current growth state.

Reverie.mp4

Why Reverie

Most agents wait for human prompts.

Reverie introduces a different primitive: subconscious generation.

• It synthesizes internal impulses/goals instead of depending on continuous human input.
• It writes memory continuously and compounds context over time.
• It can be switched between autonomous growth and interactive response.

This makes Reverie suitable for experiments in long-horizon autonomous behavior, memory-driven intelligence, and self-directed agent research.

Core Modes

1) Reverie Mode (autonomous growth)

Runs unattended. The agent continuously explores, plans, executes tools, and updates memory.

2) Lucid Mode (interactive response)

When the user asks a question, the agent responds based on the memory/state accumulated so far.

3) Murmur Mode (hybrid)

Blends both worlds: autonomous drifting + user-driven interaction.

What makes it different

• Subconscious-first loop: replaces prompt-only triggering with internally generated intent.
• Persistent memory stack: keeps state across sessions and decisions.
• Tool-native execution: search, file I/O, command execution, memory operations.
• Built for observability: API + WebUI to inspect runtime state.

Architecture at a glance

• reverie/subconscious/: subconscious prompting and state shaping
• reverie/memory/: memory types, scoring, persistence
• reverie/scheduler/: autonomous turn scheduling and queueing
• reverie/orchestrator/: planning, tool loops, execution flow
• reverie/api/: FastAPI services for dashboard/runtime access
• webui/: visualization interface for activity and memory

Quick Start (Local)

1. Install dependencies

python3 -m pip install openai python-dotenv fastapi uvicorn

2. Configure env

cp .env.example .env

Minimum required variables:

• OPENAI_API_KEY
• OPENAI_BASE_URL
• REVERIE_MODEL

3. Run by mode

python3 -m reverie.cli --mode Reverie --session-id reverie
python3 -m reverie.cli --mode Lucid --session-id lucid
python3 -m reverie.cli --mode Murmur --session-id murmur

Quick Start (Docker)

Start backend + webui:

docker compose up -d --build backend webui

Start and enter Lucid session directly:

./start.sh

Or use mode commands:

./reverie.sh reverie
./reverie.sh lucid
./reverie.sh murmur

Useful operations:

./reverie.sh logs
./reverie.sh down

WebUI: http://localhost:5173
API Health: http://localhost:8000/health

Use Cases

• Autonomous research companion that keeps generating and refining ideas
• Persistent coding assistant that accumulates project context
• Sandbox for studying emergent behavior in long-running agents
• Prototype platform for “always-on” personal AI systems

Roadmap

• Stronger long-term memory retrieval and consolidation
• Better subconscious intent quality controls
• Multi-agent collaboration and shared memory spaces
• Richer observability and evaluation benchmarks

Star This Project

If the idea of agents that grow before you ask resonates with you, please give Reverie a star.

It helps the project attract contributors and accelerate research toward truly self-evolving agents.

License

See LICENSE.