MCP memory server for Claude Code, Cursor, and any MCP client. Pure Go. Single binary. No external services required.

Claude Code's built-in memory is a markdown file with a 200-line cap. No search. No categories. No importance ranking. Every project is siloed. After ten sessions the file is 30% redundant, and the architecture decision you saved last week gets silently truncated because it landed on line 201.

Ghost replaces that with a real memory system:

One command migrates your existing Claude Code memories into Ghost. Nothing is lost.

go install github.com/wcatz/ghost/cmd/ghost@latest

This puts the binary in your $GOBIN (default ~/go/bin/). Make sure it's on your $PATH. Or download a pre-built binary from Releases and put it wherever you want.

ghost mcp init

[1/7] Checking prerequisites...
  ✓ ghost binary at ~/go/bin/ghost
  ✓ claude CLI at ~/.local/bin/claude

[2/7] Registering MCP server...
  ✓ ghost MCP server registered

[3/7] Adding tool permissions...
  + 16 mcp__ghost__* tools added to allow list

[4/7] Configuring SessionStart hook...
  + ghost hook session-start — injects project context at startup

[5/7] Disabling Claude Code built-in memory...
  + set autoMemoryEnabled: false (disables competing file-memory)

[6/7] Importing Claude Code memories...
  ✓ myproject — 8 memories imported
  ✓ infra — 12 memories imported

[7/7] Writing project memory redirects...
  ✓ myproject — redirect written
  ✓ infra — redirect written

Done! Restart Claude Code to activate.

After setup, Claude automatically loads your project context at session start and saves discoveries during work. No manual prompting needed.

ghost mcp status          # verify integration health
ghost mcp init --dry-run  # preview changes without writing

Idempotent and non-destructive — safe to re-run after updates. Existing Claude Code MEMORY.md files with user content are never overwritten. Permissions and hooks are added, never removed. Use --dry-run to preview before committing.

Add Ghost to your MCP config:

{
  "mcpServers": {
    "ghost": {
      "type": "stdio",
      "command": "ghost",
      "args": ["mcp"]
    }
  }
}

Ghost is a memory pipeline with four stages:

Save → Search → Consolidate → Decay

Save — Claude (or you) saves memories via MCP tools. Each memory has a category, importance score (0.0-1.0), and tags. FTS-based upsert deduplicates on save — if a similar memory already exists in the same category, it strengthens instead of creating a duplicate.

Search — FTS5 full-text search with optional vector similarity (Ollama embeddings). Cross-project search finds knowledge from other repos. Global memories (_global) are included in every project's context.

Consolidate — Periodic reflection merges duplicates, prunes noise, and promotes cross-project knowledge to global scope. Two tiers:

A quality gate rejects garbage output (< 30% of input) and falls through to the next tier.

Decay — Time-based scoring fades stale memories by category:

Ghost exposes 16 tools to any MCP client:

Task statuses: pending → active → done. Tasks can also be set to blocked from any non-done state and return to active when unblocked. done tasks are conventionally closed — prefer creating a new task over re-opening a completed one.

Resources (4):

Learned context is a prose summary of the project generated by ghost reflect. It captures patterns and decisions that individual memories may not convey. Ghost injects it alongside top memories in every session; it is only updated during a manual ghost reflect run, never during normal MCP usage.

The MCP server ships with embedded instructions that teach Claude when to save, which categories to use, and how to leverage cross-project search — so it works proactively without configuration.

ghost mcp                    # Run MCP server on stdio (used by Claude Code)
ghost mcp init [--dry-run]   # Configure Claude Code integration
ghost mcp status             # Check integration health
ghost hook session-start     # SessionStart hook (called by Claude Code)
ghost reflect <project>      # Manual memory consolidation (dry-run by default)
ghost upgrade                # Self-update from GitHub Releases
ghost version                # Print version

The ghost hook session-start command reads {"cwd": "<path>"} from stdin and writes a Markdown system-reminder to stdout containing top memories, open tasks, active decisions, and global memories — injected into the session before any MCP tools are available.

ghost reflect flags: --apply to save, --restore to undo, --tier haiku|sqlite|auto.

Requires Go 1.25+.

go install github.com/wcatz/ghost/cmd/ghost@latest

Or build from source:

git clone https://github.com/wcatz/ghost.git && cd ghost && make build

Download from GitHub Releases — linux, macOS, Windows (amd64 + arm64).

ghost upgrade    # self-update from GitHub Releases

docker run -v ghost-data:/data ghcr.io/wcatz/ghost:latest

Superpowers is a skills framework for AI coding agents — it enforces brainstorm-first planning, mandatory TDD, and subagent-driven execution. Ghost and Superpowers are built to complement each other: Superpowers structures how work gets done, Ghost remembers what was learned.

Session start
  └── Ghost SessionStart hook → injects project context (memories, tasks, decisions)
  
User asks for a new feature
  └── Superpowers brainstorm skill → clarifies requirements
  └── Superpowers calls ghost_project_context → loads codebase history and past decisions
  └── Superpowers writes a plan informed by Ghost's memory of conventions and gotchas

Subagents execute the plan
  └── Each phase ends with ghost_memory_save → persists what was learned
  └── Architectural choices go through ghost_decision_record
  └── Bugs found along the way → category: gotcha, importance: 0.9

Next session
  └── Ghost hook fires → top memories already in context
  └── No re-explaining the codebase, conventions, or past decisions

In any Claude Code session, run:

/plugin install superpowers@claude-plugins-official

No additional configuration needed — skills trigger automatically based on what you ask for.

Config loads in layers (later overrides earlier):

1. Compiled defaults
2. /etc/ghost/config.yaml
3. ~/.config/ghost/config.yaml
4. .ghost/config.yaml (per-project)
5. .ghost/config.local.yaml (gitignored)
6. GHOST_* environment variables

Minimal config (no file needed):

Ghost stores memories in ~/.local/share/ghost/ghost.db. No configuration required for basic MCP usage.

Optional embedding (vector search):

embedding:
  enabled: true
  ollama_url: "http://localhost:11434"
  model: "nomic-embed-text:v1.5"

Requires Ollama running locally. Enables hybrid FTS5 + vector search.

Optional reflection (memory consolidation with Haiku):

export ANTHROPIC_API_KEY="sk-ant-..."
ghost reflect myproject --apply

Without an API key, Ghost falls back to the free SQLite-based Jaccard consolidator.

cmd/ghost/main.go          CLI bootstrap
internal/
  memory/                  SQLite + FTS5 + vector search + time-decay scoring
  reflection/              Tiered consolidation (Haiku → SQLite)
  mcpserver/               MCP server (stdio, 16 tools + 4 resources)
  mcpinit/                 Claude Code integration setup (init, status, hook)
  claudeimport/            Auto-import Claude Code memory files
  ai/                      Claude API client (used by reflection only)
  embedding/               Ollama async vectorization worker
  config/                  Layered YAML/env config (koanf)
  selfupdate/              Self-update from GitHub releases
  provider/                Interface contracts (MemoryStore, LLMProvider)

Apache License 2.0 — see LICENSE.