Claude Code Memory

A Claude Code plugin that gives Claude fully automatic, per-project cognitive memory — powered by local embeddings and hybrid search.

Features

• Autonomous hook loop (v1.2) — the plugin searches memory on every user prompt, before every file edit, before every Bash command, and when Claude uses Read/Grep/WebFetch. Claude doesn't need to remember to call memory_search.
• Auto-saves per turn with a prose summary (no meta-prefix boilerplate); user corrections automatically become pattern memories.
• Auto-recalls matching memories when Bash errors occur.
• Auto-merges near-duplicate memories on store (≥95% cosine similarity) and auto-creates relations on every write path.
• Relation graph grows over time — co-injected memories promote to relates_to; a periodic sweep (every 20 sessions) finds connections missed at insert time; relation weights evolve with usage.
• Type-aware decay — patterns decay slowly and boost aggressively on recall; episodics fade fast; semantic and procedural are permanent unless you delete them.
• Cluster-aware retrieval — when a memory is injected, its strongest neighbours (weight ≥ 0.5) come along.
• Compact, markdown-stripped injection format — no section headers; one line per memory with high-importance patterns starred.

How Claude uses memory

Claude still has direct access to the MCP tools (memory_store, memory_relate, memory_update, etc.) for judgement-level work — explicit "remember this" requests, overrides, or corrections to auto-saved memories.

All data stays local. Everything is scoped to the current project via the working directory.

Requirements

• Node.js >= 18
• Claude Code CLI

Installation

git clone https://github.com/d2a8k3u/claude-code-memory.git claude-memory
cd claude-memory
./install.sh

The installer builds the MCP server, registers it globally, adds hooks to ~/.claude/settings.json, and symlinks skills. Restart Claude Code after installation.

Statusline: The installer adds a 🧠 N loaded badge (memories loaded into context this session, over the project corpus size) only if you have no statusLine configured — it never overwrites an existing one. To add the badge to your own statusline, run it yourself and append the output:

node /path/to/claude-memory/server/dist/cli.js statusline

Note: On first use, the plugin downloads a ~90 MB embedding model (all-MiniLM-L6-v2) from Hugging Face. This happens once and is cached locally. The first session start may take 10-30 seconds depending on your connection.

First-run bootstrap (optional)

Run /memory-init in any project to populate the memory from existing project files (README, package.json, CLAUDE.md, git history, etc.).

How It Works

Session Lifecycle

1. SessionStart — clears working memories, decays stale importance, injects relevant context via adaptive budget allocation
2. During session — Claude uses 10 MCP tools automatically (store, search, get, update, delete, list, batch-store, relate, graph, health)
3. PostToolUse — on Bash errors, extracts error terms and surfaces matching memories
4. Stop — analyzes the session transcript with noise filtering and saves structured memories (episodic, procedural, semantic, pattern)

Memory Types

Memories can be linked with relations: relates_to, depends_on, contradicts, extends, implements, derived_from.

Search & Scoring

Hybrid search combining FTS5 full-text and semantic vectors (cosine similarity via sqlite-vec). Falls back to text-only when embeddings are unavailable. Results are reranked with a cross-encoder (ms-marco-TinyBERT-L-2-v2) for accurate relevance scoring.

Scoring combines text relevance, importance, recency, and access frequency with configurable per-channel weight presets (e.g. branch queries favor recency, project-level queries favor importance). Verbose memories receive a content-length penalty to keep results focused.

Storage

SQLite database at .claude/memory-db/memory.sqlite inside each project. Includes FTS5 and vec0 virtual tables for fast search.

Skills & Commands

claude-memory/
├── install.sh                    # Installer
├── server/src/
│   ├── index.ts                  # MCP server entry point
│   ├── cli.ts                    # Hook runner entry point
│   ├── database.ts               # SQLite + FTS5 + vec0
│   ├── memory.ts                 # 10 MCP tool handlers
│   ├── embeddings.ts             # Local embeddings (all-MiniLM-L6-v2, 384-dim)
│   ├── thresholds.ts             # Centralized similarity/scoring configuration
│   └── cli/
│       ├── session-start.ts      # Budget allocation, context injection
│       ├── session-end.ts        # Transcript analysis, memory creation
│       ├── error-context.ts      # Error pattern matching
│       ├── pattern-detector.ts   # Shared pattern clustering logic
│       └── transcript.ts         # Transcript parsing with noise filtering
├── skills/
│   ├── memory-init/              # /memory-init bootstrap skill
│   └── memory-maintain/          # /memory-maintain cleanup skill
├── commands/
│   ├── memory-graph.md           # /memory-graph slash command
│   ├── memory-status.md          # /memory-status slash command
│   └── memory-recall.md          # /memory-recall slash command
├── agents/memory-curator.md      # Maintenance sub-agent
└── hooks/hooks.json.template     # Reference hook config

Tech Stack

TypeScript · MCP SDK · better-sqlite3 · sqlite-vec · @huggingface/transformers · Zod · tsup

Contributing

See CONTRIBUTING.md.

License

MIT