CCCMemory MCP (xiaolai) — Summary

CCCMemory MCP is a TypeScript/Node.js MCP server that provides comprehensive persistent memory for AI coding agents by indexing Claude Code (and Codex) conversation history into a SQLite database with hybrid semantic + full-text search. Originally published as claude-conversation-memory-mcp, renamed to cccmemory in v1.8.0. It ships 40+ MCP tools organized into eight categories: indexing, search, context, project management, working memory, session handoff, context injection, and tag management. The v2.0 smart chunking, RRF hybrid search (vector + FTS), and memory quality tracking make it the most full-featured SQLite-backed memory system in this batch. It is cross-tool portable (Claude Code, Claude Desktop, Codex) and installs with a single npm install -g cccmemory. Compared to seeds, it is closest to ccmemory (patrickkidd) but substitutes Neo4j+Ollama for SQLite+Transformers.js, sacrificing graph relationship modeling for simpler installation and broader tool compatibility.

CCCMemory MCP (xiaolai) — Overview

Origin

Created by xiaolai, a prominent Chinese software developer and educator. Published in 2025, renamed from claude-conversation-memory-mcp to cccmemory in v1.8.0. Active development with v2.0 major upgrade.

Philosophy

"Give Claude long-term memory by indexing conversation history with semantic search, decision tracking, and cross-project search." The design principle is comprehensiveness over simplicity — unlike minimal file-based approaches, CCCMemory provides a full-featured memory infrastructure with quality tracking, TTL support, versioned tag management, and session handoff protocols.

v2.0 Architecture Improvements

• Smart Chunking — Long messages split at sentence boundaries; previously truncated at 512 tokens
• Hybrid Search — Combines semantic search with full-text search using Reciprocal Rank Fusion (RRF)
• Dynamic Thresholds — Similarity thresholds adjust based on query length
• Extraction Validation — Reduces false positives in decision/mistake detection
• Improved Snippets — Search results highlight matching terms in context

Multi-Tool Portability

Explicit support for:

• Claude Code: ~/.claude.json config
• Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json
• Codex CLI: ~/.codex/config.toml with codex mcp add cccmemory command

Storage Modes

Two modes selectable via env var:

• Global mode (default): ~/.cccmemory.db — single database for all projects
• Per-project mode: ~/.claude/projects/<project>/.cccmemory.db — isolated per project

CCCMemory MCP (xiaolai) — Architecture

Distribution

npm package (cccmemory). Install globally.

Install

npm install -g cccmemory
cccmemory --version

Add to Claude Code config (~/.claude.json):

{
  "mcpServers": {
    "cccmemory": { "command": "cccmemory" }
  }
}

Directory Structure

Source layout (src/):

src/
├── ConversationMemory.ts    # Core memory class
├── mcp-server.ts            # MCP server entry point
├── index.ts                 # CLI entry
├── cache/                   # Caching layer
├── chunking/                # Smart chunking
├── cli/                     # CLI subcommands
├── context/                 # Context injection
├── documentation/           # Doc generation
├── embeddings/              # Vector embedding (Transformers.js)
├── handoff/                 # Session handoff
├── memory/                  # Working memory
├── parsers/                 # Conversation file parsing
├── realtime/                # Real-time indexing
├── search/                  # Hybrid search (FTS + vector)
├── storage/                 # SQLite storage
├── tools/                   # MCP tool definitions + handlers
├── types/                   # TypeScript types
└── utils/                   # Utilities

State Storage

• Primary: ~/.cccmemory.db (SQLite, single-file by default)
• Per-project: ~/.claude/projects/<project>/.cccmemory.db
• Config: ~/.claude-memory-config.json

Required Runtime

• Node.js 20 or 22 LTS (strictly — other versions break better-sqlite3)

Embedding Backend

Default: @xenova/transformers (Transformers.js) — bundled, works offline, model auto-downloaded on first use. Alternatives via config: Ollama, OpenAI.

Target AI Tools

Claude Code, Claude Desktop, Codex CLI, any MCP-compatible client.

CCCMemory MCP (xiaolai) — Components

MCP Tools (~45 total, organized by category)

Indexing (2)

Search (10)

Context (8)

Project Management (4)

Working Memory (5)

Session Handoff (3)

Context Injection (2)

Tag Management (7)

Hooks / Commands / Skills

None shipped — pure MCP server. No Claude Code plugin hooks.

CLI (via cccmemory binary)

The npm package installs a cccmemory binary. Used for starting the server; also supports --version flag.

CCCMemory MCP (xiaolai) — Uniqueness

Differs From Seeds

Closest to ccmemory (patrickkidd, seed) in purpose but uses SQLite+Transformers.js hybrid search instead of Neo4j+vector+graph. The architectural tradeoff: ccmemory's graph relationships (SUPERSEDES, CONFLICTS_WITH, DEPENDS_ON) enable causal tracing; cccmemory's hybrid RRF search (vector + FTS5) enables better recall across large conversation histories. cccmemory has 40+ tools vs ccmemory's ~10; cccmemory installs with one npm command vs ccmemory's Docker stack. Also similar to claude-flow (seed) in being MCP-anchored, but claude-flow is a full orchestration framework while cccmemory is pure memory infrastructure.

Distinctive Features

1. Widest tool coverage in this batch: 40+ MCP tools covering indexing, search, context injection, working memory, session handoff, tag management, documentation generation, project migration.
2. Cross-platform portability: Explicit support for Claude Code, Claude Desktop, Codex CLI — the only system in this batch explicitly configured for all three.
3. Smart chunking with RRF: v2.0's hybrid vector+FTS approach with Reciprocal Rank Fusion is more sophisticated than any other file-based memory system in this batch.
4. Tag management as first-class feature: Full tag lifecycle (list, search, rename, merge, delete) across all memory types.

Observable Failure Modes

1. Node version sensitivity: Requires strictly Node 20 or 22 LTS; better-sqlite3 breaks on other versions.
2. No lifecycle hooks: Context is not automatically injected at session start; agent must call get_startup_context explicitly.
3. No graph relationships: Unlike ccmemory, decisions are flat; no SUPERSEDES or CONFLICTS_WITH — stale decisions coexist with newer ones.
4. Unbounded growth: No compaction or archiving mechanism; database grows indefinitely.
5. First-run model download: Transformers.js downloads embedding model on first use, which can fail in air-gapped or sandboxed environments.