Agent Belief Infrastructure
Provegate
Three MCP servers that give AI coding agents justified, revisable beliefs about your codebase — with drift detection, persistent memory, and verified change gating.
View on GitHubQuick StartThree Layers
Layer 1
claude-drift
Detects divergence between architectural intent and implementation. Reads rules from CLAUDE.md or ADRs, derives enforceable constraints, and scores drift continuously.
Layer 2
claude-memory-mesh
Persists what agents believe — with evidence, confidence, scope, and decay. Beliefs are linked, revisable, and contradiction-aware across sessions.
Layer 3
claude-proof
Verifies changes before they enter memory. Gates what agents can commit to their belief store, ensuring only checked invariants propagate.
The Problem It Solves
- Repeated reasoning — agents rediscover the same constraint every session because nothing persists across them.
- Contradictory decisions — isolated agents form incompatible models of the same code with no reconciliation layer.
- Invisible architectural drift — code diverges from intent and no tool measures it continuously.
- Flat memory — CLAUDE.md and Cursor rules have no relationships, no evidence, no scoped validity, no decay.
Quick Start
pip install fastmcp
# Add to Claude Code (adjust paths)
claude mcp add claude-drift -- \
python /path/to/claude-drift/server.py
claude mcp add claude-memory-mesh -- \
python /path/to/claude-memory-mesh/server.py
claude mcp add claude-proof -- \
python /path/to/claude-proof/server.py