Governing AI Context with a Memory Invocation and Context Archive (MICA) Schema
This article discusses the challenges of session loss in long-running AI projects and proposes a governance schema called MICA to manage AI context and memory across sessions.
Why it matters
Effective context management is critical for the reliability and long-term viability of AI systems, especially in mission-critical applications.
Key Points
- 1Session loss is a structural problem for long-running AI projects, not just an inconvenience
- 2Existing approaches like longer prompts and session summaries treat context as a document, not a governed structure
- 3MICA is a specification-level solution that structures context with authority levels, eviction rules, and provenance
- 4MICA focuses on governing what context is allowed to shape the AI session, not just handling the output
Details
The article introduces key terms like MICA (Memory Invocation & Context Archive), Session Loss, Trust Class, Invoke Role, Invocation, and Admission Gate. It explains how v0.0.1 of the MICA schema proved that context could be structured, but failed to govern what context is allowed to shape the AI session. The author then discusses how the problem requires a different framing - treating the LLM as an unreliable upstream service and focusing on what goes in (context admission) rather than just what comes out (output validation). The MICA schema aims to define the rules for context item admission, trust level assignment, provenance checking, and eviction priority.
No comments yet
Be the first to comment