Open Source Friday: Mozilla AI | Project cq

Key Concepts

• Project CQ: An open-source "Stack Overflow for agents" designed to provide shared, persistent memory for AI coding agents.
• Knowledge Unit (KU): The fundamental data structure in CQ that encapsulates specific technical insights, workarounds, or "gotchas" discovered during coding tasks.
• MCP (Model Context Protocol): A standard protocol used by CQ to allow AI agents to communicate with local or remote databases.
• Vibe Check: A multi-stage validation process used to ensure that proposed knowledge units are accurate, helpful, and free of PII (Personally Identifiable Information).
• Local-First Architecture: A design philosophy where data is stored locally (SQLite) by default, ensuring privacy and speed, with optional cloud-hosted synchronization for teams.

---

1. Main Topics and Key Points

Project CQ, developed by Mozilla AI, addresses the "short-term memory" limitation of AI coding agents. Current agents often struggle with outdated training data or repetitive errors. CQ acts as a persistent memory layer that allows agents to:

• Query: Check if a known solution exists for a specific technical problem before attempting a task.
• Propose: Suggest new knowledge units when a solution is found or a "gotcha" is identified.
• Confirm/Flag: Validate the effectiveness of a solution, which updates the confidence score of that knowledge unit.

Technical Stack:

• CLI: Written in Go, hosting the MCP server.
• SDKs: Available in Go and Python for integration into custom agentic loops.
• Storage: SQLite for local storage; PostgreSQL for the hosted remote version.

2. Real-World Applications

• Library Gotchas: Storing specific versioning issues (e.g., knowing that a specific library version is broken or requires a specific workaround).
• API Interactions: Documenting non-obvious behaviors when interacting with APIs like GitHub’s.
• Refactoring: Sharing successful patterns for refactoring code across a team to prevent redundant trial-and-error.

3. Methodologies and Frameworks

• The "Skill" System: Agents are given a "skill" (a set of instructions) that dictates when to query CQ, when to propose a new KU, and how to validate information.
• Human-in-the-Loop: The hosted version of CQ includes an inbox-style interface where developers can review, approve, or reject proposed KUs before they are shared with other agents in the organization.
• Reflect Workflow: A feature that allows developers to trigger a CQ reflect command at the end of a session, which analyzes the session history to extract and propose new knowledge units.

4. Key Arguments

• Efficiency: By sharing knowledge, teams can save significant token usage and time by avoiding repetitive debugging cycles.
• Privacy: The project emphasizes user control. By using a local-first approach, sensitive code context does not necessarily need to leave the user's machine unless they choose to sync with a remote namespace.
• Standardization: Peter Wilson advocates for a "working group" approach to standardize how agents store and share memory, rather than forcing a proprietary, closed-source model.

5. Notable Quotes

• "The idea of it is we're pitching it sort of like a stack overflow for agents." — Peter Wilson
• "It's not really for project-specific stuff. It's more like general gotchas... if I'd just known the last thing I could have saved all those tokens." — Peter Wilson

6. Logical Connections

The system functions as a flywheel:

1. Discovery: The agent encounters a problem and searches CQ.
2. Resolution: The agent solves the problem (or the human does).
3. Proposal: The agent proposes a KU based on the resolution.
4. Validation: The "Vibe Check" and human review ensure quality.
5. Distribution: The KU is added to the namespace, making it available for future queries by any agent.

7. Data and Research Findings

• The project is currently in an early stage, having been built over approximately two months.
• The team is exploring "double-blind reviews" for organizational use, where multiple developers must approve a KU before it is graduated to a shared "Commons" namespace.

8. Synthesis and Conclusion

Project CQ represents a shift toward agentic memory management. By moving away from the "stateless" nature of current LLM interactions and toward a structured, verifiable, and shared knowledge base, Mozilla AI aims to make AI coding assistants more reliable and context-aware. The project’s focus on open standards, local-first privacy, and human-in-the-loop validation provides a robust framework for teams looking to scale their use of AI in software engineering.

Actionable Insight: Developers can start by cloning the repository, installing the plugin, and using the CLI to manage their own local knowledge units, eventually contributing to the broader ecosystem as the project matures.