Squish - Two-Tier Memory System for Claude Code

A persistent, context-aware memory plugin for Claude Code (v0.7.2) that provides intelligent memory management with a two-tier architecture: always-in-context core memory (2KB) plus context paging for selective memory loading.

Dual-Mode Operation:

• MCP Server for Claude Code (default, no args)
• CLI Commands for OpenClaw/bash execution

What's New in v0.7.2:

• CLI Mode: Execute squish remember, search, recall directly from terminal
• Local Embeddings: Zero-dependency TF-IDF embeddings (works offline, no Ollama needed)
• OpenClaw Ready: Install via npm, use via bash commands

Features

Memory Architecture

• Core Memory (Tier 1): 2KB always-visible persistent memory divided into 4 sections

  • persona: AI assistant personality and role definition
  • user_info: Information about the user and their preferences
  • project_context: Current project architecture and patterns
  • working_notes: Transient notes during active work sessions

• Context Paging (Tier 2): Agent-controlled memory loading with token budgeting

  • Dynamically load/evict memories based on relevance
  • Token-aware context management (8KB default budget)
  • Automatic memory decay and lifecycle management

MCP Tools

Core Memory Management:

• /core_memory - View, edit, or append to always-in-context memory sections

Context & Search:

• /context_paging - Load/evict/view loaded memories in current session
• /context - Get project context with semantically relevant memories
• /search - Full-text + semantic search across memories
• /context_status - Check context window usage and loaded memories

Memory Operations:

• /remember - Store new observations, facts, decisions, and preferences
• /recall - Retrieve specific memories by ID with metadata
• /observe - Record tool usage and patterns observed during execution

Memory Management:

• /merge - Detect duplicates, preview/approve/reject merges

System:

• /health - Check MCP server and database connectivity

Storage

• SQLite (local mode): Local-first, zero setup required
• PostgreSQL (team mode): Shared memory across users
• Smart embeddings: Local TF-IDF (offline) for SQLite, OpenAI/Ollama for PostgreSQL

OpenClaw Gateway (v0.7.0)

Squish provides a WebSocket gateway for OpenClaw AI assistant integration, enabling OpenClaw agents to access Squish's memory system.

Gateway Details:

• Protocol: WebSocket
• Port: 18789
• Messages: search, remember, recall, status

Enable Gateway:

export OPENCLAW_GATEWAY_ENABLED=true

Message Format:

{
  "type": "search",
  "id": "unique-request-id",
  "payload": {
    "query": "authentication flow",
    "limit": 10
  }
}

Response Format:

{
  "type": "search",
  "id": "unique-request-id",
  "success": true,
  "data": {
    "results": [...],
    "count": 10
  }
}

Installation

Install via Claude Code Marketplace (Recommended)

# Step 1: Add the Squish marketplace
/plugin marketplace add https://github.com/michielhdoteth/squish.git

# Step 2: Install Squish
/plugin install squish@michielhdoteth-squish

For Local Development Testing

Load the plugin directly without installation:

claude --plugin-dir "C:\Users\michi\Desktop\squish-cc\squish"

Or on macOS/Linux:

claude --plugin-dir "/path/to/squish"

This loads the plugin from your development directory and allows immediate testing of changes.

For Production Installation

1. Build the plugin:

npm install
npm run build

2. Package as npm module:

npm pack

3. Install the .tgz file:

npm install -g squish-memory-0.7.0.tgz

Configuration

Environment Variables

Required (SQLite mode - default):

• None! Works out-of-the-box with local TF-IDF embeddings

Optional (SQLite mode):

• SQUISH_DATA_DIR: Custom data directory (default: ./.squish in project root)
• SQUISH_EMBEDDINGS_PROVIDER: Force embedding provider (none, local, openai, ollama)

Team Mode (PostgreSQL):

• DATABASE_URL: PostgreSQL connection string (activates team mode)
• SQUISH_OPENAI_API_KEY: OpenAI API key for embeddings (optional in team mode)
• SQUISH_OLLAMA_URL: Ollama server URL for local LLM embeddings

OpenClaw Gateway (v0.7.0):

# Enable OpenClaw WebSocket gateway
OPENCLAW_GATEWAY_ENABLED=true

Advanced Options:

# Lifecycle management
SQUISH_LIFECYCLE_ENABLED=true
SQUISH_LIFECYCLE_INTERVAL=3600000  # ms

# Session summarization
SQUISH_SUMMARIZATION_ENABLED=true
SQUISH_INCREMENTAL_THRESHOLD=10
SQUISH_ROLLING_WINDOW_SIZE=50

# Agent isolation
SQUISH_AGENT_ISOLATION_ENABLED=true
SQUISH_DEFAULT_VISIBILITY=private  # private|project|team|global

Quick Start

CLI Usage (OpenClaw/Bash)

Install and use directly from terminal:

npm install -g squish-memory

# Store memories
squish remember "User prefers TypeScript for all projects" --type preference
squish remember "API uses JWT authentication" --type fact

# Search memories
squish search "authentication" --limit 5
squish search "TypeScript" --type fact

# View core memory
squish core_memory view

# Retrieve by ID
squish recall <memory-id>

Memory Types: fact, decision, preference, observation, context

Embedding Options:

By default, Squish uses local TF-IDF embeddings (zero dependencies, works offline).

For better semantic search, configure an embedding provider:

# OpenAI embeddings (best quality)
export SQUISH_OPENAI_API_KEY=sk-...
export SQUISH_EMBEDDINGS_PROVIDER=openai

# Ollama local embeddings (good quality, runs locally)
export SQUISH_EMBEDDINGS_PROVIDER=ollama
export SQUISH_OLLAMA_URL=http://localhost:11434

# Or set in your shell profile (~/.bashrc or ~/.zshrc)
echo "export SQUISH_OPENAI_API_KEY=sk-..." >> ~/.bashrc
echo "export SQUISH_EMBEDDINGS_PROVIDER=openai" >> ~/.bashrc

MCP Server (Claude Code)

1. Initialize Core Memory

Store basic information that will always be available:

/core_memory action=edit projectId=my-project section=persona content="I am a helpful assistant specialized in cloud infrastructure and DevOps."

2. Store Session Observations

Record what you learn during work:

/remember projectId=my-project type=observation action=discovered target=kubernetes_pattern summary="Found the team uses GitOps with ArgoCD" tags=["kubernetes","gitops","devops"]

3. Search Your Memory

Find relevant past learnings:

/search projectId=my-project query="kubernetes deployment" limit=5

4. Load Context for an Agent

Load relevant memories for agent use:

/context_paging action=load sessionId=session-123 projectId=my-project memoryId=mem-456

Architecture

Database Schema

Core Tables:

• core_memory - Always-in-context 2KB persistent memory
• memories - Main memory storage with lifecycle tracking
• context_sessions - Per-session context window tracking
• conversations - Chat session records
• messages - Message history with embeddings

Lifecycle Tables:

• memory_associations - Semantic relationships between memories
• memory_merge_proposals - Duplicate detection proposals
• memory_merge_history - Audit trail of merges
• memory_snapshots - Before/after diffs for auditability
• session_summaries - Incremental and rolling summaries

Search & Discovery:

• entities - Named entities (functions, files, patterns, people)
• entity_relations - Relationships between entities
• memory_hash_cache - Hash signatures for duplicate detection

Memory Lifecycle States

Each memory has:

• Sector: episodic, semantic, procedural, autobiographical, working
• Tier: hot (recent/relevant), warm (accessible), cold (archived)
• Status: active, merged, superseded, expired
• Visibility: private (user), project, team, global

Hooks

4 lifecycle hooks automatically trigger:

API Reference

Core Memory Operations

# View all sections
/core_memory action=view projectId=PROJECT_ID

# Edit a section
/core_memory action=edit projectId=PROJECT_ID section=persona content="..."

# Append to a section
/core_memory action=append projectId=PROJECT_ID section=working_notes content="New observation"

Memory Management

# Store a memory
/remember projectId=PROJECT_ID type=fact content="..." tags=["tag1","tag2"]

# Search memories
/search projectId=PROJECT_ID query="search term" type=fact limit=10

# Retrieve by ID
/recall memoryId=MEMORY_ID

# Get merged view
/context projectId=PROJECT_ID include=memories include=observations limit=20

Context Paging

# Load memory into session context
/context_paging action=load sessionId=SESSION_ID projectId=PROJECT_ID memoryId=MEMORY_ID

# Unload memory from session
/context_paging action=evict sessionId=SESSION_ID memoryId=MEMORY_ID

# View loaded memories
/context_paging action=view sessionId=SESSION_ID

# Check context usage
/context_status projectId=PROJECT_ID sessionId=SESSION_ID

Memory Merging

# Detect duplicate memories
/merge projectId=PROJECT_ID action=detect

# List merge proposals
/merge projectId=PROJECT_ID action=list status=pending

# Preview a merge
/merge projectId=PROJECT_ID action=preview proposalId=PROPOSAL_ID

# Approve merge
/merge projectId=PROJECT_ID action=approve proposalId=PROPOSAL_ID reviewNotes="Approved because..."

# Reject merge
/merge projectId=PROJECT_ID action=reject proposalId=PROPOSAL_ID reviewNotes="Keep separate because..."

# Get merge statistics
/merge projectId=PROJECT_ID action=stats

Database Modes

SQLite Mode (Default - Local Development)

• Storage: ~/.squish/squish.db (local SQLite database)
• Embeddings: Local TF-IDF (no API calls, works offline)
• Performance: Fast, suitable for single-user development
• Setup: Zero configuration required
• Data: Stored locally on machine

PostgreSQL Mode (Team - Enterprise)

• Storage: Shared PostgreSQL database
• Embeddings: Optional OpenAI or Ollama for better semantic search
• Performance: Optimized for multi-user teams
• Setup: Requires DATABASE_URL configuration
• Data: Shared across team members with access control

# Enable PostgreSQL mode
DATABASE_URL=postgresql://user:password@host/dbname
SQUISH_OPENAI_API_KEY=sk-... # optional

Benchmarks

Squish has been tested against industry-standard datasets for memory systems.

LongMemEval Dataset

Key Strengths

• Strong accuracy: 80% on LongMemEval benchmark
• Flexible deployment: Local embeddings with quality generation
• Privacy-first: Local embeddings option available
• Consistent retrieval: Successfully finds entities like names, locations, and relationships

Test Coverage

The benchmark tests memory retrieval across:

• Long-term conversational memory
• Entity resolution (people, places, organizations)
• Temporal reasoning (events, dates, timelines)
• Relationship tracking (family, professional, social)

Running Benchmarks

cd ../squish-benchmarks

# Pull Nomic embeddings
ollama pull nomic-embed-text

# Run LongMemEval benchmark
bun run src/index.ts run -e nomic-embed-text -b longmemeval

For detailed benchmark results and methodology, see squish-benchmarks.

Development

Project Structure

squish/
├── .claude-plugin/
│   ├── plugin.json          # Plugin manifest
│   └── marketplace.json     # Marketplace metadata
├── .mcp.json                # MCP server configuration
├── api/
│   └── openclaw/
│       └── gateway.ts       # OpenClaw WebSocket gateway (v0.7.0)
├── core/
│   ├── core-memory.ts       # Core memory operations
│   ├── context-paging.ts    # Memory loading system
│   ├── local-embeddings.ts  # TF-IDF embeddings
│   ├── embeddings.ts        # Embedding providers
│   └── logger.ts
├── db/
│   ├── index.ts             # Database adapter
│   ├── bootstrap.ts         # Schema initialization
│   └── adapter.ts           # SQLite/PostgreSQL abstraction
├── drizzle/
│   ├── schema.ts            # PostgreSQL schema
│   └── schema-sqlite.ts     # SQLite schema
├── skills/
│   └── squish-memory/
│       └── SKILL.md         # OpenClaw skill definition (v0.7.0)
├── commands/                # Slash commands
├── hooks/                   # Lifecycle hooks
├── index.ts                 # MCP server entry point
├── config.ts                # Configuration management
├── package.json
├── tsconfig.json
└── README.md

Building

# Install dependencies
npm install

# Build TypeScript
npm run build

# Run tests
npm run test

# Pack for distribution
npm pack

Testing

# Run comprehensive system test
node test-v0.5.0.mjs

# Start development with watch mode
npm run dev

Troubleshooting

Database Issues

Error: "database disk image is malformed"

• SQLite database corrupted - delete ~/.squish/squish.db and restart

Error: "Connection refused" (PostgreSQL mode)

• Verify DATABASE_URL is correct: postgresql://user:pass@host:5432/dbname
• Ensure PostgreSQL server is running

MCP Server Issues

Error: "Hook error: ERR_UNSUPPORTED_ESM_URL_SCHEME"

• Use relative paths in hooks.json instead of absolute Windows paths
• Use ./hooks/file.js instead of C:\path\to\hooks.js

Error: "Plugin not found in marketplace"

• Using --plugin-dir flag? This is expected - not installed from marketplace
• For local development, this is normal behavior

Performance Issues

Memory is growing unbounded

• Enable lifecycle management: SQUISH_LIFECYCLE_ENABLED=true
• Review memory decay rates in configuration

OpenClaw Gateway Issues (v0.7.0)

Gateway not starting on port 18789

• Check if port is already in use: lsof -i :18789 (macOS/Linux) or netstat -ano | findstr :18789 (Windows)
• Verify gateway is enabled: echo $OPENCLAW_GATEWAY_ENABLED
• Check logs for connection errors

OpenClaw client cannot connect

• Ensure gateway is running: Check Squish logs for "OpenClaw gateway listening on port 18789"
• Verify firewall allows connections on port 18789
• Check WebSocket client is using correct protocol (ws:// or wss://)

Contributing

1. Fork the repository
2. Create a feature branch: git checkout -b feature/my-feature
3. Commit changes: git commit -am 'Add feature'
4. Push to branch: git push origin feature/my-feature
5. Submit a pull request

License

MIT - See LICENSE file for details

Support

• GitHub Issues: https://github.com/michielhdoteth/squish/issues
• Documentation: https://github.com/michielhdoteth/squish#readme
• OpenClaw: https://github.com/openclaw/openclaw - Personal AI assistant platform

Changelog

v0.7.0 (Latest)

• OpenClaw Gateway: WebSocket API for OpenClaw AI assistant integration
• Enhanced Search: Improved semantic search with better ranking
• Memory Tools: Expanded toolkit for memory management
• OpenClaw Skill: Ready-to-use skill for OpenClaw installation

v0.6.0

• Implemented hybrid memory retrieval with multi-factor scoring (semantic, recency, coactivation, importance)
• Added fact extraction for structured knowledge discovery
• Added entity resolver for tracking and deduplicating entities across sessions
• Refactored capture.ts for better type safety
• Updated plugin-wrapper to use smart search config (50 candidates → top 3)
• Added utility functions for hook error handling
• Reorganized test scripts into scripts/db/ and scripts/test/
• Updated .gitignore to exclude references/ and opencode/
• Cleaned up project for distribution

v0.5.0

• Implemented smart embeddings strategy (local TF-IDF for SQLite, OpenAI/Ollama for PostgreSQL)
• Added missing core_memory and context_sessions tables to bootstrap
• Fixed plugin manifest validation for Claude Code compatibility
• Fixed hooks path resolution for Windows compatibility
• Consolidated 18 MCP tools to 11 with action-based API
• Core memory fully functional with 4-section architecture

v0.4.1

• Consolidated 25 command files to 15
• Removed 4 phantom commands
• Fixed validation order in MCP request handlers

v0.3.0

• Lifecycle management with memory decay
• Session summarization (incremental and rolling)
• Agent-aware memory isolation
• Memory governance and protection
• Memory associations and graph traversal

v0.2.0

• Vector embeddings for semantic search
• Memory snapshots for auditability
• Folder-scoped observations

v0.1.0

• Initial release with core memory and context paging
• Full-text search (FTS5)
• Two-tier architecture