Name: AgentRecall
Author: goldentrii

Long-term Memory for Agents

AgentRecall provides a robust memory layer that allows agents to 'inhale' context at the start of a session and 'exhale' learned insights at the end. It moves beyond simple storage by implementing a correction-first model, tracking where agents were wrong to prevent repeating mistakes.

Core Features:

The Two-Verb Model: Simplified session_start and session_end flow for maximum efficiency.
Memory Palace: Organized long-term knowledge rooms with salience scoring.
Awareness System: Cross-project insight matching to apply lessons from one project to another.
Correction Tracking: Predictive warnings (watch_for) based on past human corrections.
Local & Private: Zero cloud, zero telemetry, and fully Obsidian-compatible markdown storage.
Semantic Search: Optional Supabase backend for vector-based cosine similarity recall.