Soul-OS

Multi-Agent Cognitive Ecosystem

A constellation of AI agents, each with distinct personalities and expertise, sharing memory and collaborating naturally with you and each other.

Currently building: Foundation

The Vision

Building a multi-agent cognitive ecosystem where different AI personalities collaborate with shared memory, creating a more natural and powerful collaborative experience.

Shared Memory

Agents share knowledge through global, project, and agent-specific memory banks, building on each other's work.

Distinct Personalities

Each agent has unique expertise, conversational style, and problem-solving approach, creating natural collaboration.

Intelligent Routing

System automatically routes tasks to the right agents or enables multi-agent collaboration for complex problems.

Current Architecture

Live

Soul-OS Frontend

→Cloudflare Worker: soul-os-frontend

→D1 Database: memory, sessions, traces

→MindBridge API: api.soul-os.cc

→OpenAI: 8 models (gpt-4o, o1, etc.)

→Anthropic: 10 models (claude-sonnet-4-6, etc.)

Cognitive Pipeline: 8-stage biomimetic processing (INGEST → WERNICKE → RAG → THALAMUS → INSULA → ASSEMBLE → GENERATE → CONSOLIDATE)

The Constellation

Six milestones to build the multi-agent ecosystem

Milestone 1•Week 1

Foundation

Model Switching & Identity

In Progress

Give each model a persistent identity and let users switch between them seamlessly within conversations.

Features

→Add `model` parameter to chat endpoint
→Create agent profiles with personality and expertise
→Store which agent responded to each message
→Simple UI for agent selection

Deliverables

Agent profiles database tableUpdated chat endpoint with model selection3-5 default agents (Architect, Builder, Researcher, etc.)Agent attribution in message history

Technical Implementation Notes

• Extend existing D1 schema with `agents` table
• Add `agent_id` field to messages table
• Update model adapter to accept agent profiles
• Create agent selection component

Milestone 2•Week 2-3

Shared Memory Banks

Multi-Scope Memory

Pending

Agents share memory but can also have private knowledge. Memory scopes enable both collaboration and specialization.

Features

→Memory scopes: global, agent-specific, project
→Agents can read global memory (shared knowledge)
→Agents can write to their own memory bank
→Memory attribution tracking

Deliverables

Memory scope field in databaseAgent-specific memory filteringGlobal memory consolidationMemory attribution UI

Technical Implementation Notes

• Add `scope` and `owner_agent_id` to memory_items table
• Update RAG search to filter by scope
• Create memory visibility rules
• Implement cross-agent memory access patterns

Milestone 3•Week 4

Agent Awareness

@Mentions & Handoffs

Pending

Agents can reference and invoke each other through natural @mention syntax, enabling collaborative problem-solving.

Features

→@mention syntax for agent invocation
→Agents can see conversation history from other agents
→Simple agent-to-agent communication
→Handoff pattern: one agent defers to another

Deliverables

Mention parser in message processingAgent-to-agent message routingConversation context sharingHandoff workflow implementation

Technical Implementation Notes

• Parse @mentions in user messages
• Route messages to mentioned agents
• Include other agents' messages in context
• Create handoff protocol

Milestone 4•Week 5-6

Automatic Routing

Intelligent Orchestration

Pending

System intelligently routes to the right agent(s) based on task type, enabling seamless multi-agent collaboration.

Features

→Wernicke router enhancement for agent selection
→Auto-invoke relevant agents based on task
→Multi-agent responses (optional)
→Consensus mode for important decisions

Deliverables

Task classification systemAgent routing logicMulti-agent response aggregationConsensus algorithm

Technical Implementation Notes

• Extend Wernicke router with agent selection
• Create task-to-agent mapping
• Implement parallel agent invocation
• Design consensus mechanism

Milestone 5•Week 7-8

Relational Dynamics

Personality & Rapport

Pending

Agents have distinct personalities and build relationships with users, creating a more natural collaborative experience.

Features

→Personality traits in system prompts
→Agents remember user preferences
→Agents develop rapport (stored in memory)
→Agents can disagree, debate, or defer
→User-configurable agent personalities

Deliverables

Personality system prompt templatesUser preference trackingRapport scoring systemAgent interaction patternsPersonality configuration UI

Technical Implementation Notes

• Create personality trait system
• Track user-agent interactions
• Implement rapport scoring
• Design agent debate protocols

Milestone 6•Week 9-10

Advanced Memory

Learning & Evolution

Pending

Agents learn from interactions and evolve over time, developing expertise and adapting to user preferences.

Features

→Agent-specific skill development tracking
→Cross-session learning
→Memory consolidation per agent
→Agent expertise growth metrics

Deliverables

Skill tracking systemLearning metrics dashboardAgent evolution visualizationExpertise growth algorithms

Technical Implementation Notes

• Create skill development schema
• Implement learning algorithms
• Build expertise tracking
• Design evolution metrics