Goldentrii/AgentRecall-X

What it solves

AgentRecall provides a long-term memory system for AI agents that specifically focuses on learning from human corrections. While most memory tools focus on simply retrieving information, AgentRecall tracks whether a correction actually changes the agent's behavior across different sessions, preventing the agent from repeating the same mistakes.

How it works

The system implements a "learning loop" consisting of a governed corrections ledger and a measurement harness. It organizes memory into five distinct layers: Episodic (session journals), Semantic (topic-clustered facts), Procedural (IF-THEN rules), Narrative (project phases), and Correction (behavioral calibration rules).

It integrates with agents via an MCP server, SDK, or CLI, using a session-based workflow: /arstart to load context, /arrecall to search for knowledge, /arsave to save and compound learning, and /arreflect to consolidate memory. Retrieval is handled through a combination of keyword matching and Reciprocal Rank Fusion (RRF), with optional vector search support.

Who it’s for

Developers using AI agents (specifically Claude Code, Cursor, Windsurf, or VS Code) who want their agents to persistently remember user preferences and technical corrections across projects and sessions.

Highlights

• Behavioral Measurement: Tracks "heed rates" and "phantom gradient steps" to quantify if corrections actually stop mistakes.
• Multi-Layer Memory: Uses a cognitive-psychology inspired taxonomy to organize data into journals, rooms, skills, and awareness.
• Local-First: Stores all memory as local markdown files by default, ensuring privacy and zero cloud dependency.
• Automated Consolidation: Includes an optional "Dreaming" mode for nightly autonomous pattern mining and memory decay management.
• War Room Dashboard: A local-first visual interface to monitor agent learning and project status.