context-memory

Persistent, searchable context storage across Claude Code sessions using SQLite + FTS5.

Table of Contents

• Why?
• Features
• Architecture
• Installation
• Uninstalling
• Requirements
• Commands
• How It Works
• Usage Examples
• Trigger Phrases
• Web Dashboard
• CLI Tools
• Database Management
• Testing
• Contributing
• License
• Author

Why?

Claude Code sessions are ephemeral - every conversation starts from zero. Close the terminal and everything you discussed, decided, and solved is gone. The only way to get that context back is to re-explain it or hope Claude reads the right files.

context-memory fixes this.

• "How did we fix that auth bug last week?" - Instead of re-debugging, /recall authentication pulls up exactly what you did, what you decided, and why.
• Decisions have context - You chose JWT over sessions for a reason. Three months later, you can't remember why. Your past sessions can.
• Code patterns survive sessions - That elegant retry pattern you built? It's saved with the language, file path, and description of what it does. /recall retry pattern --detailed brings it back.
• Projects you haven't touched in months - /recall --project scopes to whatever you're working in. Instant refresher on where you left off.
• Cross-project learning - Solved a CORS issue in one project? When it hits another, /recall CORS finds it regardless of which project it came from.
• Sub-50ms search - It's SQLite with FTS5 and tuned PRAGMAs (WAL mode, 64MB cache, in-memory temp store). Searching thousands of sessions feels instant.
• Two words to save everything - /remember and Claude does the rest: summarizes, extracts topics, identifies key code, stores it all. Add a note if you want, or don't.

Without it, every session is a blank slate. With it, Claude Code has a long-term memory that grows more valuable the more you use it.

Features

Core:

• Cross-session memory - Save and recall past work across Claude Code sessions
• Structured AI summaries - /remember generates rich summaries with brief/detailed text, key decisions, problems solved, technologies used, and outcome classification (success/partial/abandoned)
• Full-text search - FTS5 with Porter stemming for fast, fuzzy search
• Two-tier retrieval - Summary-ranked search with multi-source boost (<10ms) + deep content fetch (<50ms)
• Project-scoped or global - Filter by current project or search everything
• Topic categorization - Auto-extracted topics for browsable history
• Code snippet storage - Preserve important code with language and context
• Outcome tracking - Every session is classified as success, partial, or abandoned — searchable and charted in the dashboard

Hooks:

• Auto-save on exit - Stop hook captures session context automatically when Claude Code exits, with smart head+tail transcript sampling (first 5 + last 10 messages) to preserve the problem statement and resolution
• Git branch capture - Auto-save detects the current git branch and adds it as a topic, so /recall feature/auth-refactor finds sessions from that branch
• Pre-compact checkpoints - Saves full conversation before context compaction for zero context loss, recoverable via the context_load_checkpoint MCP tool
• Session deduplication - Auto-save checks for a rich /remember save in the same project within a configurable window (default 5 minutes) and skips if one exists
• Loop prevention - Auto-save checks stop_hook_active to prevent recursive hook invocation

Extras:

• Web dashboard - Full SPA with 17 REST API endpoints, dark/light theme, Chart.js analytics, Highlight.js code rendering, session CRUD, and search autocomplete
• MCP server - Six tools for programmatic access from any MCP-compatible client
• CLI tools - All core scripts (db_save.py, db_search.py, db_prune.py, db_init.py) have full argparse CLIs with --help
• Database pruning - Prune old sessions by age or count, and old checkpoints per session, with dry-run preview

Engineering:

• Cross-platform - Windows (CMD/PowerShell), macOS, and Linux
• Zero external dependencies - Stdlib-only Python 3.8+ for core functionality
• Schema auto-migration - Forward-only automatic upgrades (v1 through v4) on first DB access after upgrade, preserving all existing data
• SQLite performance tuning - WAL mode, 64MB cache, PRAGMA synchronous=NORMAL, PRAGMA temp_store=MEMORY
• 364 tests across 12 modules - CI runs on Python 3.8, 3.11, and 3.12 with ruff linting

Architecture

.claude-plugin/plugin.json        # Plugin manifest (version, metadata)
.mcp.json                         # MCP server config (project-level)
install.py                        # Installer (idempotent, selective flags)
uninstall.py                      # Uninstaller (copies itself to survive clone deletion)
hooks/hooks.json                  # Hook definitions (Stop + PreCompact)
commands/                         # /remember and /recall command definitions
skills/context-memory/            # Skill definition (SKILL.md)
  scripts/
    __init__.py                   # Package init, version, public API re-exports
    db_init.py                    # Schema creation, verification, stats, migrations
    db_save.py                    # Session storage logic, deduplication
    db_search.py                  # FTS5 search (tier 1 + tier 2)
    db_prune.py                   # Database pruning (by age/count/checkpoints)
    db_utils.py                   # Connection management, helpers, shared utilities
    mcp_server.py                 # MCP server (FastMCP, stdio transport)
    dashboard.py                  # Web dashboard (Flask REST API + SPA)
    auto_save.py                  # Stop hook: cross-platform auto-save wrapper
    pre_compact_save.py           # PreCompact hook: saves full context before compaction
    static/                       # Dashboard frontend (~2,400 lines vanilla JS/CSS/HTML)
      index.html                  # SPA entry point (Chart.js, Highlight.js CDN)
      dashboard.css               # Dual-theme CSS with custom properties
      js/app.js                   # Router, theme toggle, navigation
      js/api.js                   # REST API client
      js/components/              # Charts, code blocks, modals, toasts, session cards
      js/views/                   # Search, sessions, detail, analytics, settings
  references/
    schema-reference.md           # Full database schema reference
tests/                            # 364 tests across 12 modules
.github/workflows/ci.yml          # CI: lint (ruff) + test (Python 3.8, 3.11, 3.12)

Installation

git clone https://github.com/ErebusEnigma/context-memory.git
cd context-memory
python install.py

The installer copies the skill, commands, and hooks (Stop + PreCompact) to the correct Claude Code locations (~/.claude/) and initializes the database. It's idempotent — run it again to upgrade. On upgrade, the installer detects outdated hooks (old command paths, unexpanded ~ on Windows) and updates them in-place.

After installation, the cloned directory is no longer needed and can be deleted (or kept for future upgrades). The uninstaller is copied to ~/.claude/context-memory/uninstall.py so it works even after deleting the clone.

cd .. && rm -rf context-memory   # optional cleanup

Installer flags:

Note: The hooks use Python wrappers (auto_save.py, pre_compact_save.py) and work cross-platform — Windows (CMD, PowerShell), macOS, and Linux.

Uninstalling

python ~/.claude/context-memory/uninstall.py

This removes the skill, commands, hooks (both Stop and PreCompact), and MCP server registration. Your saved sessions are preserved by default. Use --remove-data to also delete the database, or --keep-data to skip the prompt. Use --force to remove command files even if they've been modified.

Requirements

• Python >= 3.8
• SQLite with FTS5 support (included in Python's standard library)
• MCP server (optional): Python >= 3.10 and pip install mcp
• Web dashboard (optional): pip install flask flask-cors

Commands

/remember [note]

Save the current session to context memory.

/remember
/remember "Fixed the auth bug with refresh tokens"
/remember "Important: OAuth2 implementation details"

Claude will automatically:

1. Generate a structured summary (brief + detailed text, key decisions, problems solved, technologies used)
2. Classify the session outcome (success, partial, or abandoned)
3. Extract topics and identify important code snippets
4. Store everything in the local SQLite database

/recall <query> [options]