Calmkeep

Calmkeep addresses a specific challenge in AI-assisted legal work: the tendency of large language models to drift from established positions, assumptions, or reasoning frameworks over the course of extended sessions. The product functions as an external oversight layer, operating alongside the model to detect and document inconsistencies before they affect downstream legal work.

Key Features and Functions:

Session-Level Continuity Monitoring
Tracks legal frameworks, jurisdictional selections, and strategic positions established earlier in a session. Identifies and flags deviations as they occur across extended AI interactions.

Structural Drift Detection
Detects changes in legal reasoning or assumptions that may not be explicitly acknowledged by the model. Focuses on maintaining consistency in areas such as governing law, legal theories, and cited authorities.

External Enforcement Layer
Operates independently of the underlying model without requiring fine-tuning or retraining. Integrates at the deployment level to provide real-time oversight during AI-assisted workflows.

Audit and Evaluation Framework
Provides structured evaluation of AI session outputs against predefined criteria. Includes documentation of deviations, scoring methodologies, and session transcripts for review.

Integration and Deployment Options
Available through a Python SDK, MCP connector, and Claude Code plugin. Supports use with existing Anthropic API access through a bring-your-own-key (BYOK) model.

Data Handling Approach
Designed to operate without server-side data retention. Processes session data in real time to support continuity monitoring.