Runtime memory engine
Contorium is AI runtime infrastructure—not a chat UI. A local pipeline turns workspace activity into evolving runtime state your agents can use across Codex, Claude Code, and IDE workflows.
Pipeline
From raw editor and Git events to agent-ready context.
Workspace scanner
Continuously observes editors, saves, focus changes, and Git events. Respects ignore rules and repo size budgets so large monorepos stay usable.
Memory builder
Ranks active files, infers goals, attaches Git context, and writes structured blocks—not raw file dumps—for downstream compression and export.
Memory compression
Semantic summaries, ranked paths, and token-aware exports keep long-running AI workflows within budget without losing what matters.
Export & restore
One action produces markdown or JSON context for agents—or restores session state, editors, and memory blocks after a restart.
Local-first: .contora/
All state lives in your workspace. No hidden cloud memory store—you control retention, sharing, and optional BYOK analysis keys.
<workspace-root>/ ├── .contoraignore └── .contora/ ├── state.json # extension snapshot (MCP reads) ├── events/ ├── last-intent.json ├── memory/ └── mcp/ └── memories.json # MCP store_memory
Design principles
- AI-triggered retrieval — agents pull memory via MCP or export, not manual paste
- Cross-session restore — memory survives chat switches and IDE restarts
- Git-aware summaries — staged and modified paths drive ranking
- Agent export formats — markdown, json, cursor, claude, openai
- Standalone runtime — future headless mode beyond the editor
See it in action
Connect agents via MCP or install the extension to start building workspace memory.