Junto Memory: A Persistent Shared Memory System for Multi-Agent AI Workflows

Junto Memory is an MCP (Model Context Protocol) server designed for Multi-Agent AI workflows, focusing on solving three key pain points in multi-agent collaboration: memory gaps, cross-agent coordination issues, and context degradation. The system uses a dual-storage architecture of MongoDB + ChromaDB, supporting long-term collaboration of multi-agent teams. It has been validated in production environments, coordinating 6 professional agents to complete the development of a commercial IoT platform with over 500 sessions, and includes a complete capability matrix of more than 40 tools across 14 categories.

There are three core issues in multi-agent collaboration: 1. Memory gaps: Knowledge disappears after an agent session ends, requiring subsequent agents to re-learn; 2. Collaboration conflicts: Multiple agents modify the same file simultaneously, lacking resource locking and state awareness; 3. Context degradation: Long sessions lead to decreased model performance even before reaching the capacity limit. Junto Memory is designed specifically to address these problems.

Junto Memory uses a dual-storage backend of MongoDB + ChromaDB: MongoDB is responsible for storing documents, states, tasks, and messages, while ChromaDB provides vector search capabilities. Unlike most MCP memory servers that only support single agents, it is designed for multi-agent teams and supports long-term project collaboration over weeks or months. Its 40+ tools cover 14 categories, with key capabilities including multi-agent session tracking, cross-agent message delivery, file locking, function registry, task to-do management, etc. (most typical MCP servers do not support these features).

The first interaction of an agent usually follows this flow:

1. memory_start_session: Start a session to get the session ID, previous learning records, active work, etc.;
2. memory_record_learning: Store learning content (e.g., technical notes), which will be automatically returned in subsequent sessions;
3. memory_register_function: Register function information, which future agents can find via semantic search. Three calls are enough to establish persistent memory and avoid blank starts.

Junto Memory runs via a Docker Compose stack, and AI agents connect via MCP (streamable HTTP transmission on port 8080). All knowledge, messages, and states are persisted in MongoDB and ChromaDB volumes, so no data is lost upon restart. The installation process is suitable for AI agent operations: after cloning the repository, the agent can independently complete steps such as prerequisite checks, environment configuration, and container startup.

Junto Memory has been validated in heterogeneous technology stacks such as C#/.NET servers, Raspberry Pi Python, and .NET MAUI mobile applications, with all solved problems derived from actual production experience. In the MCP ecosystem, it fills the gap in team-level multi-agent coordination, promoting AI agents from individual operations to organized team collaboration.

As AI agents play an increasingly important role in software development, managing agent collaboration, memory, and context has become a key challenge. Junto Memory provides a production-validated solution that not only solves technical problems but also establishes a new paradigm: agents are no longer isolated individuals but team members with a shared brain, continuous learning, and collaboration capabilities.