Contextrie is a long-term memory framework for AI agents designed to prevent context rot across extended workflows. Instead of appending endless chat history or relying on rigid RAG pipelines, it treats context as a curated artifact. The system is organized around a three-stage pipeline: Ingest, Assess, Compose.
Ingest loads source files such as markdown, text, and CSV, then generates metadata for each item, including a high-compression summary and atomic keypoints that act as relevance hooks. Assess scores each source against a task from 0.0 to 1.0 in either shallow mode (metadata only) or deep mode (full content analysis), so the agent can weigh what matters now. Compose then compresses and formats the selected sources into final context with density presets, preserving the most relevant material at higher fidelity and compressing the rest.
The goal is to keep agents sharp from day one to day 1000 by continually selecting the right information at the right time. The project is early but functional, with planned work around organic memory, assembly loops, per-item assessment, and broader file support. Contextrie focuses on clarity, relevance, and controlled compression so agents can act decisively without drowning in their own history.