MegaMemory

MegaMemory

Persistent project knowledge graph for coding agents.

An MCP server that lets your coding agent build and query a graph of concepts, architecture, and decisions — so it remembers across sessions.

The LLM is the indexer. No AST parsing. No static analysis. Your agent reads code, writes concepts in its own words, and queries them before future tasks. The graph stores concepts — features, modules, patterns, decisions — not code symbols.

The Loop

understand → work → update

1. Session start — agent calls list_roots to orient itself
2. Before a task — agent calls understand with a natural language query (or get_concept for exact ID lookup)
3. After a task — agent calls create_concept or update_concept to record what it built

Everything persists in a per-project SQLite database at .megamemory/knowledge.db.

Installation

npm install -g megamemory

[!NOTE] Requires Node.js >= 18. The embedding model (~23MB) downloads automatically on first use.

Quick Start

megamemory install

Run the interactive installer and choose your editor:

[!NOTE] The installer only updates config files after a successful read/merge, and it will not overwrite existing plugin/command files unless they are already marked as MegaMemory-managed.

With opencode

megamemory install --target opencode

One command configures:

• MCP server in ~/.config/opencode/opencode.json
• Workflow instructions in ~/.config/opencode/AGENTS.md
• Skill tool plugin at ~/.config/opencode/tool/megamemory.ts
• Bootstrap command /user:bootstrap-memory for initial graph population
• Save command /user:save-memory to persist session knowledge

Restart opencode after running install.

With Claude Code

megamemory install --target claudecode

Configures:

• MCP server in ~/.claude.json
• Workflow instructions in ~/.claude/CLAUDE.md
• Commands in ~/.claude/commands/

With Antigravity

megamemory install --target antigravity

Configures:

• MCP server in ./mcp_config.json (workspace-level)

With Codex

megamemory install --target codex

Configures:

• MCP server in ~/.codex/config.toml
• Workflow instructions in ~/.codex/AGENTS.md

With other MCP clients

Add megamemory as a stdio MCP server. The command is just megamemory (no arguments). It reads/writes .megamemory/knowledge.db relative to the working directory, or set MEGAMEMORY_DB_PATH to override.

{
  "megamemory": {
    "type": "local",
    "command": ["megamemory"],
    "enabled": true
  }
}

MCP Tools

Concept kinds: feature · module · pattern · config · decision · component

Relationship types: connects_to · depends_on · implements · calls · configured_by

Knowledge Graph

Web Explorer

Visualize the knowledge graph in your browser:

megamemory serve

• Nodes are colored by kind and sized by edge count
• Dashed edges show parent-child links; solid edges show relationships
• Click any node to inspect summary, files, and edges
• Search supports highlight/dim filtering
• If port 4321 is taken, you'll be prompted to pick another

megamemory serve --port 8080   # custom port

How It Works

src/
  index.ts       CLI entry + MCP server (9 tools)
  tools.ts       Tool handlers (understand, get_concept, create, update, link, remove, list_conflicts, resolve_conflict)
  db.ts          SQLite persistence (libsql, WAL mode, schema v3)
  embeddings.ts  In-process embeddings (all-MiniLM-L6-v2, 384 dims)
  merge.ts       Two-way merge engine for knowledge.db files
  merge-cli.ts   CLI handlers for merge, conflicts, resolve commands
  types.ts       TypeScript types
  cli-utils.ts   Colored output + interactive prompts
  install.ts     multi-target installer (opencode, Claude Code, Antigravity, Codex)
  web.ts         HTTP server for graph explorer
plugin/
  megamemory.ts  Opencode skill tool plugin
commands/
  bootstrap-memory.md  /user command for initial population
  save-memory.md       /user command to save session knowledge
web/
  index.html     Single-file graph visualization (d3-force + Canvas)

• Embeddings — In-process via Xenova/all-MiniLM-L6-v2 (ONNX, quantized). No API keys, and no network calls after the first model download.
• Storage — SQLite with WAL mode, soft-delete history, and schema migrations (currently v3).
• Search — Brute-force cosine similarity over node embeddings; fast enough for graphs with <10k nodes.
• Merge — Two-way merge with conflict detection by concept ID, with AI-assisted conflict resolution via MCP tools.

CLI Commands

Merging Knowledge Graphs

When branches diverge, each may update .megamemory/knowledge.db independently. Since SQLite files cannot be auto-merged by git, megamemory provides dedicated merge commands.

Merge two databases

megamemory merge main.db feature.db --into merged.db

Concepts are compared by ID: identical nodes are deduplicated, and conflicting nodes are kept as ::left/::right variants under one merge group UUID. Use --left-label and --right-label to replace default side labels with branch names.

megamemory merge main.db feature.db --into merged.db --left-label main --right-label feature-xyz

View conflicts

megamemory conflicts            # human-readable summary
megamemory conflicts --json     # machine-readable output
megamemory conflicts --db path  # specify database path

Resolve conflicts manually

megamemory resolve  --keep left    # keep the left version
megamemory resolve  --keep right   # keep the right version
megamemory resolve  --keep both    # keep both as separate concepts

AI-assisted resolution

When an AI agent runs /merge, it calls list_conflicts, verifies both versions against current source files, then calls resolve_conflict with resolved: {summary, why?, file_refs?} plus a verification reason. It does not pick a side blindly; it resolves to what the codebase currently reflects.

License

MIT