About this project
Magic Context is a plugin for coding agents that provides self-managing context and long-term memory, described as the hippocampus for coding agents and part of the CortexKit family. It aims to eliminate compaction pauses and memory loss across sessions by capturing durable knowledge, consolidating memories overnight, and recalling relevant information automatically.
The project supports OpenCode, Pi, and Oh My Pi (OMP) harnesses. An interactive setup wizard detects installed harnesses, configures models, disables built-in compaction, and resolves conflicts with other context-management plugins. Manual configuration is also documented.
Core features include historian compartmentalization that compresses old history into tiered chronological summaries, decay rendering that adjusts fidelity deterministically, cache-stable layout to preserve prompt cache, and optional agent-driven reduction. Memory capture extracts decisions, constraints, conventions, and config values into project memory. A dreamer agent can run overnight to verify memories against the codebase, curate duplicates, classify importance, maintain documentation, and build user profiles. Recall tools include ctx_search across memories, conversation history, git commits, notes, and primers; ctx_expand to restore compressed transcripts; and ctx_note for deferred intentions.
A companion desktop app provides memory browsing, session history, cache diagnostics, dreamer management, configuration editing, and log viewing. State is stored in a local SQLite database under the shared CortexKit store. The project is MIT licensed and distributed via npm packages.
Comments
0 people shared their preference · Deer Point appears after 10 participants
Sign in to join the discussion.