MemoryStack for Claude Code

Give Claude Code a brain that persists. MemoryStack is a cognitive memory engine for AI agents — and this plugin brings it to Claude Code. Your agent remembers what you worked on, how you code, and what decisions you made — across every session.

Why MemoryStack?

AI coding agents start every session with amnesia. They forget your preferences, your architecture decisions, the bugs you've already fixed. MemoryStack fixes this with cognitive memory — not just storage, but intelligent recall:

• 🧠 Importance Scoring — Automatically ranks memories by significance
• 🔍 Hybrid Search — Semantic similarity + keyword matching for precise recall
• 🔄 Auto-Consolidation — Deduplicates and merges related memories
• 📉 Temporal Decay — Recent memories surface first, old ones gracefully fade
• 💡 Reflection — Periodically synthesizes insights from accumulated knowledge
• 🤝 Multi-Agent — Agents share memories within teams, or keep them private

What This Plugin Does

This plugin hooks into 6 points in Claude Code's lifecycle to build a complete memory layer:

| When | What Happens | |------|-------------| | Session Start | Injects your Developer Profile, Project Knowledge, and Recent Activity | | Every Prompt | Detects memory signals ("remember this...") and searches for relevant context | | Every Tool Use | Tracks file edits, commands, and code changes in real-time | | Subagent Finishes | Captures what subagents accomplished for future reference | | Task Completes | Auto-saves completed team tasks as project knowledge | | Session End | Extracts key decisions, patterns, and learnings from the conversation |

The Result

<memorystack-context>
The following is recalled context about the developer and this project.

## Developer Profile
- Prefers TypeScript with strict mode, uses Bun as package manager
- Uses conventional commits with scope prefixes
- Prefers functional patterns over class-based OOP

## Project Knowledge
- Auth uses JWT with refresh tokens stored in httpOnly cookies
- API follows REST conventions with /api/v1 prefix
- Database is PostgreSQL with Drizzle ORM

## Recent Activity
- Refactored auth middleware to support multi-tenant (2h ago)
- Fixed CORS issue with WebSocket connections (1d ago)
</memorystack-context>

Quick Start

1. Get Your API Key

Visit memorystack.app/dashboard to create a free account and generate an API key.

2. Install

# Set your API key
export MEMORYSTACK_API_KEY="ms_proj_..."

# Install from npm and add to Claude Code
claude plugin add @memorystack/claude-code

Or install from source:

git clone https://github.com/memorystack-labs/memorystack-claude-code.git
cd memorystack-claude-code
npm install && npm run build
claude plugin add ./plugin

3. Use Claude Code Normally

That's it. Memories are captured and recalled automatically. You can also:

• Save explicitly: Tell Claude "remember this architecture decision"
• Search memories: Ask "what did I work on yesterday?"
• Index a project: Run /memorystack-claude-code:index to snapshot your codebase

Features

🔮 Smart Context Injection

On session start, the plugin fetches a structured profile — not a flat list of memories, but organized categories:

• Developer Profile — Your coding preferences, habits, and conventions
• Project Knowledge — Architecture decisions, patterns, team conventions
• Recent Activity — What you worked on recently, with relative timestamps

📡 Real-Time Signal Detection

The plugin listens to every prompt you type. When it detects signal phrases like "remember this", "architecture decision", or "important convention", it flags the turn for priority capture.

🔧 Tool Activity Tracking

Every file edit, bash command, and code change is compressed into a one-liner and logged. This builds a rich activity feed that the session summary uses to create focused, useful memories instead of noisy transcripts.

🤖 Subagent Memory

When Claude spawns subagents (Explore, Plan, Bash), the plugin reads each subagent's separate transcript and saves a compressed summary. Subagent work doesn't disappear — it becomes searchable project knowledge.

🏗 Dual-Scope Memories

Memories are scoped to either you (personal) or your project (shared):

| Scope | What Gets Stored | Who Sees It | |-------|------------------|-------------| | Personal | Your preferences, habits, coding style | Only you | | Project | Architecture, decisions, conventions | Anyone on the project |

Search with scope flags: --personal, --project, or --both (default).

📋 Team Task Tracking

When using Claude Code's agent teams, completed tasks are automatically saved as project knowledge with full context: who completed the task, what it was, and when.

Commands

/memorystack-claude-code:index

Index your codebase into MemoryStack. Explores project structure, architecture, conventions, and saves comprehensive findings.

/memorystack-claude-code:logout

Clear saved MemoryStack credentials.

Configuration

Environment Variables

MEMORYSTACK_API_KEY=ms_proj_...   # Required — your API key
MEMORYSTACK_DEBUG=true             # Optional — enable debug logging

Settings File

Create ~/.memorystack-claude/settings.json for advanced configuration:

{
  "captureMode": "smart",
  "signalKeywords": ["remember", "decision", "convention", "important", "bug", "architecture"],
  "turnsBefore": 2,
  "maxContextResults": 5,
  "debug": false,
  "skipTools": ["Read", "Glob", "Grep"],
  "captureTools": ["Edit", "Write", "Bash", "Task"]
}

| Setting | Default | Description | |---------|---------|-------------| | captureMode | "smart" | "smart" (signal → full fallback), "signal", or "full" | | signalKeywords | See above | Keywords that trigger priority capture | | turnsBefore | 2 | Context window: include N turns before each signal | | maxContextResults | 5 | Max memories to inject per category |

How It's Built

The plugin uses MemoryStack's SDK (@memorystack/sdk) under the hood. It's built with esbuild and ships as bundled CommonJS scripts in the plugin/scripts/ directory.

plugin/
├── .claude-plugin/plugin.json     # Plugin manifest
├── hooks/hooks.json               # 6 lifecycle hooks
├── commands/                      # /index, /logout
├── skills/                        # memory-search, memory-save
└── scripts/                       # 9 bundled .cjs scripts

Development

git clone https://github.com/memorystack-labs/memorystack-claude-code.git
cd memorystack-claude-code
npm install
npm run build       # Build all 9 scripts
npm run lint        # Check code quality

About MemoryStack

MemoryStack is a cognitive memory engine for AI agents. It goes beyond simple vector storage:

• Importance Scoring — Not all memories are equal. MemoryStack knows which ones matter.
• Contradiction Detection — When new information conflicts with old, it resolves intelligently.
• Memory Consolidation — Related memories merge over time, just like human memory.
• Temporal Awareness — Recent memories are more accessible; old ones decay naturally.
• Reflection — Periodic synthesis generates higher-order insights from raw memories.

Used by developers building AI assistants, coding agents, customer support bots, and multi-agent systems.

SDKs: Node.js · Python Docs: memorystack.app/docs

License

MIT

Support

• Website: memorystack.app
• Docs: memorystack.app/docs
• GitHub: memorystack-labs/memorystack-claude-code
• Discord: discord.gg/dHu5hMXd