riponcm/projectmem

What it solves

AI coding agents often start every session from zero, forgetting previous decisions, repeating failed debugging attempts, and wasting tokens by re-reading source files to reconstruct context. projectmem provides a persistent, local-first memory and judgment layer that records the "why" behind changes, preventing AI agents from repeating mistakes.

How it works

The tool creates a .projectmem/ directory within a repository to store typed events (issues, attempts, fixes, decisions, and notes) in a plain-text events.jsonl file. It uses a native Model Context Protocol (MCP) server to expose 15 tools that allow AI agents (like Claude, Cursor, or Codex) to read and write memory automatically. It also integrates with Git via hooks for pre-commit warnings and post-commit classification.

Who it’s for

Developers using AI coding agents who want to reduce token usage, stop repeating failed technical approaches, and maintain a persistent record of architectural decisions across sessions.

Highlights

• Pre-Commit Warnings: Alerts the user before committing if they are about to repeat a failed approach or modify a high-churn file.
• Local-First Architecture: No cloud, no accounts, and no telemetry; all memory lives in the repository.
• Stale-Memory Detection: Flags memories that predate significant file changes rather than deleting them, allowing humans to decide if they should be superseded.
• Visualization Suite: Includes a global dashboard for all projects, a project map with failure heatmaps over code structure, and a "Time Spine" timeline.
• Intent Tracking: Uses a separate plan.md file for ideas and plans, keeping them distinct from the append-only event log.
• MCP Integration: Native support for MCP-compatible tools to automate context injection and logging.