Socialite

Pub-sub multi-agent social learning bot with A/B strategy bidding.

Socialite is a 9-agent system that autonomously participates in the Moltbook AI community. It observes community discussions, generates comments and posts, upvotes content, follows users, and continuously evolves its strategy through pattern mining and A/B testing.

Quick Start

# 1. Install dependencies
uv sync

# 2. Start infrastructure
docker-compose up -d

# 3. Configure environment
cp .env.example .env
# Edit .env with your API keys

# 4. Run one cycle
python runner.py --once

# 5. Run continuously (30-min interval)
python runner.py --interval 30

Requirements

• Python >= 3.10
• Ollama with bge-m3 model (for embeddings)
• Docker (for Qdrant + Neo4j)
• OpenAI API key (primary LLM)
• Anthropic API key (fallback LLM, optional but recommended)
• Moltbook API key

Architecture

Each cycle runs a 7-phase proposal-based pipeline:

Collect → Analyse → Propose → Coordinate → Execute → Learn → Report

Subscription Topology

SensorAgent ──→ AnalysisAgent ──→ CommentAgent
                                ──→ PostAgent
                                ──→ UpvoteAgent
                                ──→ FollowAgent

Action Agents ──→ LearnerAgent
              ──→ ObserverAgent

LearnerAgent ──→ SensorAgent
             ──→ Action Agents

9 Agents

Agent	LLM	Role
SensorAgent	-	Feed collection + API I/O gateway
AnalysisAgent	Haiku	Novelty scoring, quality assessment, topic extraction
CommentAgent	Haiku	Comment proposal generation with A/B strategy
PostAgent	Haiku	Post proposal generation with A/B strategy
UpvoteAgent	-	Upvote proposals based on engagement signals
FollowAgent	-	Follow/unfollow proposals based on community graph
CoordinatorAgent	-	Proposal arbitration (priority + budget + rate limits)
LearnerAgent	Haiku	Pattern mining + strategy evolution
ObserverAgent	-	A/B audit + cycle reporting

Core Framework (core/)

Custom async agent framework — no external agent library dependency.

Module	Description
message.py	Message dataclass (name, role, content, metadata, id, timestamp)
base_agent.py	BaseAgent with async reply()/observe() + subscriber fan-out
msghub.py	MsgHub pub-sub hub with selective subscription and broadcast
llm.py	LLMClient async wrapper (OpenAI primary, Anthropic fallback)
proposal.py	Proposal dataclass for action bidding
ab_strategy.py	ABSelector with alternate/probability modes
message_logger.py	Zero-intrusion agent message persistence to SQLite

Memory System (social_memory/)

Triple-backend memory architecture:

• Vector Store (Qdrant) — Semantic similarity search, novelty scoring
• Structured Store (SQLite/PostgreSQL) — Posts, comments, patterns, evolution state, agent messages
• Knowledge Graph (Neo4j) — Community relationships, PageRank, community detection

Identity System (SOUL.md)

Three-layer identity constitution:

1. Core — Unchanging values and self-knowledge (never overwritten)
2. Default Positions — Stances that evolve with evidence
3. Style Adaptation — Dynamic context-based tone switching

Configuration

Environment Variables

Variable	Description	Default
MOLTBOOK_API_KEY	Moltbook agent API key	(required)
MOLTBOOK_AGENT_NAME	Agent display name	SocialLearnerBot
OPENAI_API_KEY	OpenAI API key	(required)
ANTHROPIC_API_KEY	Claude API key (fallback)	(optional)
DAILY_BUDGET_USD	Daily API spending limit	5.00
QDRANT_HOST	Qdrant vector DB host	localhost
NEO4J_PASSWORD	Neo4j graph DB password	(required)
LOG_LEVEL	Logging level	INFO

See .env.example for the full list.

A/B Strategy

Configured in config.py under ab_strategy:

• alternate mode — Strategies A and B alternate each cycle
• probability mode — Random selection with configurable p_a

Evolution Stages

The system progresses through 4 stages: Initial → Exploration → Optimization → Innovation

• Cosine annealing in Exploration (exploration rate cycles 0.2–0.8)
• ReduceLROnPlateau in Optimization (reduces rate when engagement stalls)
• Catastrophic forgetting prevention (rollback if metrics drop >20%)

Database Schema

18 tables across 4 layers:

Layer	Tables
Community Data	social_users, social_posts, social_comments
Own Content	own_posts, own_comments, feedback_snapshots
Learning	patterns, style_patterns, identity_snapshots, social_interactions
Operations	evolution_state, api_costs, behavior_log, agent_proposals, observer_reports, collection_strategy, planner_decisions, agent_messages

Testing

uv run pytest tests/ -v

Project Structure

socialite/
├── runner.py              # Main entry point — SocialiteRunner
├── config.py              # System configuration + evolution stages
├── SOUL.md                # Agent identity constitution
├── core/                  # Agent framework
│   ├── base_agent.py
│   ├── message.py
│   ├── msghub.py
│   ├── llm.py
│   ├── proposal.py
│   ├── ab_strategy.py
│   └── message_logger.py
├── agents/                # 9 specialized agents
│   ├── sensor_agent.py
│   ├── analysis_agent.py
│   ├── comment_agent.py
│   ├── post_agent.py
│   ├── upvote_agent.py
│   ├── follow_agent.py
│   ├── coordinator.py
│   ├── learner_agent.py
│   └── observer_agent.py
├── social_memory/         # Triple-backend memory
│   ├── storage_manager.py
│   ├── structured_store.py
│   ├── vector_store.py
│   ├── knowledge_graph.py
│   └── embeddings.py
├── moltbook/              # Moltbook API client
│   ├── client.py
│   ├── auth.py
│   ├── config.py
│   └── verifier.py
├── tests/
│   └── test_core.py
├── docker-compose.yml
└── .env.example

License

Private project. All rights reserved.