OzBrain: A Shared Knowledge Layer for Multi-Agent Workflows

OzBrain provides a unified knowledge layer that prevents context drift between AI agents

OzBrain is a hosted knowledge base designed to serve as a "shared brain" for AI agents. Instead of manually copying briefs, preferences, and project status between different LLM platforms—such as Claude, ChatGPT, and Cursor—OzBrain provides a single source of truth that all connected agents can read from and write to. This architecture ensures that every agent operates on the most current information, regardless of which tool is being used.

Core Functionality and Agent Integration

OzBrain operates as a layer beneath the AI agents, utilizing the Model Context Protocol (MCP) to integrate directly into agent environments.

Automated Knowledge Management

OzBrain moves beyond static note-taking by treating knowledge maintenance as a designed behavior for agents:

• Automated Routing: New information is automatically routed to the correct article rather than requiring the user to design a complex filing system.
• Conflict Resolution: When an agent attempts to write information that contradicts existing data, the system pauses the write and surfaces the conflict for resolution.
• Dynamic Refactoring: To prevent context window saturation and degradation, OzBrain automatically splits and reshapes articles that become too large, ensuring agents pull only the specific information needed for a task.
• Provenance Tracking: Every change is versioned and attributed to the specific agent (e.g., claude-code or chatgpt) that performed the write, providing a transparent audit log of how knowledge evolved.

Integration Flow

Users connect OzBrain to their agents via custom connector menus. For example, in Claude, the process involves adding a custom connector using the OzBrain MCP URL (https://ozbrain.com/api/mcp), signing in via an email code, and granting approval. Once connected, the agent can autonomously read the brain for context and write back learned information.

Security and Data Sovereignty

To address the privacy concerns inherent in cloud-based knowledge stores, OzBrain implements several security layers:

• Encryption at Rest: Article bodies are sealed using per-account keys. Data is decrypted only to serve agents or perform maintenance tasks like refactoring.
• Row-Level Security (RLS): The system uses forced RLS in Postgres to ensure strict tenant isolation.
• Auditability: Users have access to a full exportable CSV log of every read and write operation, including which agent accessed which article.
• Data Portability: Users can export their entire brain as plain Markdown files at any time, and account deletion results in a hard delete of all content.

Comparison with Existing Memory Systems

OzBrain distinguishes itself from platform-native memory and traditional knowledge bases in several ways:

Feature	Platform Memory (e.g., ChatGPT Memory)	Traditional Wikis (Notion/Obsidian)	OzBrain
Scope	Single platform/account	Human-centric notes	Cross-platform agent brain
Ownership	Controlled by AI provider	Controlled by human	Shared between human and agents
Maintenance	Automatic but opaque	Manual human effort	Agent-driven, structured routing
Interoperability	Locked to one ecosystem	Manual import/export	Native MCP connector

Community Perspectives and Technical Critiques

Discussion among developers on Hacker News highlights a tension between hosted agent-memory services and local-first workflows.

The Case for Local-First Workflows

Several users argued that existing tools are sufficient for their needs. One user noted that storing .md files directly in a Git repository allows agents to use grep and local file access without needing an MCP server. Others mentioned using the "Johnny.Decimal" system in Obsidian to provide agents with precise file paths, arguing that basic structure is more effective than a dedicated service.

Concerns Regarding LLM Reliability

Technical critiques focused on the inherent instability of LLM-generated summaries. One contributor questioned how OzBrain handles the degradation of accuracy when compiling large amounts of LLM-produced text, noting that even state-of-the-art models can distort meaning during non-trivial summarization batches.

The "Agent Stack" Evolution

Despite the critiques, some users viewed OzBrain as a critical piece of the emerging "agent stack." One user observed:

"It’s very encouraging to see serious attempts at addressing continuity between agents’ outputs. Right now, everyone seems to be figuring out their own way of maintaining consistency across sessions without endlessly over-contextualizing each new one."

Pricing Tiers

OzBrain offers a tiered pricing model based on the volume of knowledge (number of articles) stored:

• Free: Up to 50 articles, unlimited reads/writes.
• Pro ($20/mo): Up to 300 articles.
• Max ($99/mo): Up to 600 articles.
• Company: Custom pricing for organization-owned shared brains.