Injects last operational state into every agent session and creates backups before context compaction. Reads a checkpoint file (e.g. state/current.md) and injects it as context so the agent always knows where it left off. Warns when the checkpoint is stale. Use when building persistent agents that need continuity across sessions, context compactions, or restarts.
Semantic memory recall for AI agents using a local Qdrant vector database. Searches past conversations, matches keywords against a facts file, and routes queries to relevant knowledge files. Use when building agents that need long-term memory, fact verification, or domain-specific knowledge retrieval — all without sending data to the cloud.
Automatically captures corrections, decisions, facts, and lessons from agent conversations and stores them in a local Qdrant vector database. Filters system noise with 30+ skip patterns, deduplicates via SHA256, and only captures user messages. Use when building agents that should learn from conversations without manual memory management.
Train your agent by talking. Detects approval/disapproval signals in conversations and builds a behavioral preference database. Injects learned preferences into agent context before each response. RLHF-lite without model fine-tuning — the agent adapts its behavior based on accumulated human feedback patterns.