New System Gives AI Coding Agents Structured Memory of Project Decisions

As coding agents now drive rapid development cycles in many software teams, the institutional knowledge behind those changes often gets lost. A new system called MOOSEDev addresses this gap by equipping agents with a structured, knowledge-graph-based project memory that tracks the "why" behind code modifications.

How It Works#

MOOSEDev stores architectural decisions, learned lessons, constraints, and design rationales in a knowledge graph that agents can query through a Model Context Protocol (MCP) interface. Each record includes lifecycle status, documentation of its origin, and links to any decisions it supersedes, creating a traceable chain of reasoning. The system uses MOOSE, a proprietary neurosymbolic reasoning engine that prioritizes symbolic knowledge as its primary layer for understanding and decision-making.

Performance Against Current Tools#

Researchers benchmarked MOOSEDev against a production vector-memory system using a public dataset of 835 typed records. On tasks requiring precise understanding—such as tracking which decisions supersede others, checking whether a result set was complete, and handling negation queries—MOOSEDev substantially outperformed the baseline. The new system achieved near-perfect recall (0.98–1.00) on these logic-dependent tasks, while the vector-based alternative recovered only 6–27% of relevant answers.

For simpler relevance matching and computational cost, both systems performed comparably, suggesting that structured memory excels where logical precision matters most.

What's Next#

The researchers validated their approach through temporal analysis of their own codebase using commit history, plus a pre-registered live trial. The work also documents practical lessons from deployment, providing a foundation for integrating explicit project memory into agent-driven workflows.