Show HN: Reference MCP – let your AI agents search each other's past sessions

The recent Show HN post for "Reference MCP" addresses a friction point that has quietly plagued advanced AI users: the siloing of context between different AI sessions. The tool, built on the Model Context Protocol (MCP), allows one AI agent (e.g., Claude Code) to query the past sessions of another agent (e.g., Codex) directly, retrieving decisions, rationale, and code history without manual copy-pasting.

What Happened

The developer created an MCP server that, when prompted, establishes a session bridge between different AI instances. Instead of asking Claude to recall what Codex decided in a previous conversation—a process that relies on vague user memory or manual note-taking—the tool gives the active agent direct programmatic access to the historical session data. The repository provides the infrastructure for this cross-agent memory retrieval, effectively turning past conversations into a queryable database.

Why It Matters

This solves a specific but significant problem: the "context fragmentation" that occurs when users switch between AI tools for different stages of a project. A developer might use Codex for initial architecture decisions, then Claude Code for implementation. Without a shared memory layer, the second agent operates blind to the first agent's reasoning. Reference MCP creates a primitive but functional "institutional memory" for AI workflows.

The implications extend beyond convenience. When AI agents cannot reference each other's work, users become the bottleneck—forced to manually summarize and transfer context. This defeats the purpose of delegation. By enabling agent-to-agent memory sharing, Reference MCP moves toward a multi-agent architecture where tools collaborate autonomously. It also highlights the growing importance of the Model Context Protocol as a standard for inter-agent communication, rather than just a tool for connecting AI to external data sources.

Implications for AI Practitioners

For developers and power users, this tool suggests a new pattern: treat every AI session as a persistent, queryable artifact. Instead of viewing conversations as ephemeral, practitioners can now build workflows where later agents audit, extend, or critique earlier work. This is particularly valuable for debugging—an agent can trace why a specific decision was made by examining the original session's reasoning chain.

However, the approach raises practical concerns. Session data may contain sensitive information, and granting cross-agent access without careful scoping could leak context across projects. Additionally, the quality of retrieval depends on how well the original session was structured—rambling or unfocused conversations will yield poor references. The tool is currently a proof-of-concept, but it points toward a future where AI agents maintain a shared, searchable history, reducing the cognitive load on the human orchestrator.

Key Takeaways

• Reference MCP uses the Model Context Protocol to let AI agents query each other's past sessions, solving the problem of context fragmentation across different AI tools.
• This enables a multi-agent workflow where later agents can audit and extend the reasoning of earlier agents without manual human intervention.
• Practitioners should consider session structure and data privacy when implementing cross-agent memory, as retrieval quality depends on conversation clarity.
• The tool signals a shift toward persistent, queryable AI session histories as a standard component of advanced development workflows.