Skip to main content

Changelog

v0.1.0​

Initial release.

Added​

  • Semantic merge engine with cosine similarity deduplication
  • Similarity retrieval engine with confidence scoring and recency weighting
  • Dual mistake/success memory graphs with positive reinforcement — record_success, Reinforcement, RetrievalResponse.reinforcements
  • Graph persistence — JSON (atomic writes), SQLite, and Postgres backends (driftguard-ai[postgres])
  • MCP server with register_mistake, register_success, query_memory, deep_prune, graph_stats, guard_metrics
  • LangGraph adapter — make_langgraph_review_node
  • Generic payload adapter — review_payload
  • guard_step decorator for wrapping agent step functions
  • Guard policies: warn, block, acknowledge, record_only
  • Deep pruning engine — weak edges, stale nodes, isolated nodes (runs over both memory graphs)
  • Runtime metrics — counters and gauges for all review and storage operations
  • Offline benchmark harness — merge precision/recall and retrieval precision/recall
  • Rule-based demo agent (offline, no model downloads required)
  • LangGraph demo agent (LLM-driven planner loop)
  • Full pytest suite
  • MIT License