checkpoint

---

“Every new AI session starts from zero. Checkpoint makes it start from where you left off.”

Your AI coding assistant is brilliant within a single conversation — but start a new session and everything is forgotten. The API quirk it debugged for 30 minutes? Gone. The user preference you corrected? Forgotten. The architectural decision it finally understood? Rediscovered from scratch.

Checkpoint fixes this. One command (/checkpoint) captures every breakthrough, failed approach, user preference, and project context into structured, persistent memory files that survive across sessions — so your AI never solves the same problem twice.

The Problem

AI coding assistants are brilliant within a single conversation — but start a new session and everything is forgotten:

• Session 1: Spend 30 minutes discovering the API requires site-based paths
• Session 2: Spend 30 minutes rediscovering the same thing
• Session 3: You're frustrated. The AI has amnesia.

The Solution

Checkpoint creates a structured, persistent memory system that survives across conversations:

You:  /checkpoint
AI:   Scanning conversation... Found 3 new discoveries:

      1. [SAVED] project_auth_config.md
         - Tenant blocks app consent, must use pre-authorized client ID
         - Type: project

      2. [SAVED] feedback_no_admin.md
         - Never suggest "ask your admin" — user has no admin rights
         - Type: feedback

      3. [UPDATED] reference_api_paths.md
         - Added newly discovered endpoint paths
         - Type: reference

      Memory index updated. 3 entries saved/updated.

How It Works

Checkpoint operates on two layers:

Active Layer: /checkpoint Command

Explicitly trigger a full knowledge archival. The AI will:

1. Scan the conversation for all discoveries
2. Check existing memory to avoid duplicates
3. Consolidate — fix index issues (ghosts, orphans, cross-scope duplicates)
4. Save new findings as structured memory files (or update existing ones)
5. Update the MEMORY.md index
6. Verify completeness and report consolidation findings

Memory Consolidation: /checkpoint:consolidate

Over time, memories accumulate duplicates and stale entries. Run /checkpoint:consolidate for deep maintenance:

• Duplicate detection — find and merge memories covering the same topic (with your approval)
• Cross-scope resolution — detect duplicates between project and global memory
• Staleness detection — flag project memories with past dates or outdated references
• Quality checks — identify memories missing required structure (Why/How to apply)
• Index repair — remove dead links, add unindexed files

Passive Layer: Anti-Amnesia Protocol

A set of rules added to your CLAUDE.md (or equivalent) that runs automatically:

• Before work: Check memory for prior discoveries
• After breakthroughs: Save immediately, don't wait
• Before retries: Check if already solved — don't rediscover

Memory Types

Type	Purpose	Example
user	Who you are	"Senior backend engineer, new to React"
feedback	What you corrected	"Don't mock the database in integration tests"
project	Context behind the work	"Auth rewrite is compliance-driven, not tech debt"
reference	Where to find things	"Pipeline bugs tracked in Linear project INGEST"

Installation

Claude Code

claude install-skill Jyo238/checkpoint

This installs:

• /checkpoint slash command
• /checkpoint:consolidate sub-command for memory maintenance
• Skill definition for the AI to follow
• Reference docs for memory types, consolidation algorithm, and anti-amnesia protocol

Optional: Add the Anti-Amnesia Protocol to your project's CLAUDE.md:

cat skills/checkpoint/references/anti-amnesia-protocol.md >> CLAUDE.md

Cursor

Copy the rule file to your project:

mkdir -p .cursor/rules
cp cursor/rules/checkpoint.mdc .cursor/rules/

VSCode / GitHub Copilot