Magic Context Released to Enhance Coding Agent Memory

Magic Context, a plugin designed to solve the 'anterograde amnesia' problem where coding agents forget existing knowledge when the context window fills up or a session ends during long conversations, has been released. This plugin manages context in the background without interrupting the agent's work, maintaining persistent project knowledge.

Core Mechanisms:

• Historian: Compresses older records into hierarchical compartments and assigns importance scores in the background. It uses cheaper local models instead of the main model to improve efficiency.
• Decay Rendering: Optimizes the context window by gradually fading out older records instead of abruptly removing them.
• Cache Protection: Prevents prompt cache invalidation and reduces costs by queuing data deletions for batch processing.

Three Stages of Memory Management:

• Capture: Automatically extracts decisions, constraints, conventions, etc., and promotes them to project memory.
• Consolidate: During idle time, a 'dreamer' agent runs to cross-reference memories with the codebase, remove duplicates, and update relevant documentation.
• Recall: Automatically injects relevant memories every turn, or performs embedding-based semantic search via ctx_search.

Magic Context is written in TypeScript and supports OpenCode, Pi, Oh My Pi (OMP) hosts. All data is stored in a local SQLite database, providing persistent memory across sessions and hosts.