Waggle-mcp

Core

This repository is the public Waggle product repo: Apache-2.0 licensed, available on GitHub and PyPI, and focused on the local-first memory engine.

Related MCP server: Graphiti Knowledge Graph MCP Server

Quick Start

# Install globally (no venv needed)
pipx install waggle-mcp

# One-line setup — detects your MCP clients and writes config
waggle-mcp setup --yes

# Verify everything is healthy
waggle-mcp doctor

(No pipx? Run brew install pipx && pipx ensurepath first.)

setup --yes detects Claude Code, Codex, Cursor, Gemini CLI, and Antigravity, writes the MCP config, and installs automatic memory hooks where supported. Restart your client and you're live.

Windows users: Run all commands with python -X utf8 or set PYTHONUTF8=1 to avoid UnicodeEncodeError from emoji in log output.

Install Waggle

Waggle is a local MCP server that gives coding agents persistent graph memory.

Recommended:

• VS Code: install the live Waggle: Local Memory for AI Agents extension from the Marketplace for one-click setup

• MCP clients: use docs/install and Smithery metadata in smithery.yaml

• Claude: use docs/install/claude-code.md or docs/install/claude-desktop.md

• Developers: pipx install waggle-mcp

Benchmark:

• LongMemEval 500-case retrieval-only: 97.4% R@5, 89.0% Exact@5 for graph_raw retrieval (artifact)

VS Code extension features:

• one-click Enable for this Workspace onboarding

• installs waggle-mcp with consent if it is missing

• safely creates or updates .vscode/mcp.json

• preserves existing non-Waggle MCP servers

• runs waggle-mcp doctor

• opens Graph Studio

• exports Waggle memory from the editor

Claude distribution:

• Claude Code does not use an .mcpb bundle. Users add Waggle directly as an MCP server:

pipx install waggle-mcp
claude mcp add --transport stdio waggle -- waggle-mcp serve --transport stdio

• Claude Desktop uses the claude-desktop-extension.mcpb bundle, which can be distributed through GitHub Releases.

Manual MCP config:

{
  "mcpServers": {
    "waggle": {
      "command": "waggle-mcp",
      "args": ["serve", "--transport", "stdio"]
    }
  }
}

Enterprise Evaluation

For self-hosted production review and security posture:

• Production deployment guide

• Security model

• Hardening checklist

• Reference

60-Second Demo

No MCP client needed. Run this from a fresh install:

waggle-mcp demo

This imports a pre-loaded example graph and runs 4 scripted queries locally — no API key, no network, no client required. Add --with-embeddings to use the real sentence-transformers model for higher-fidelity retrieval (requires ~420 MB download on first run).

Why Waggle

waggle-mcp is a local-first memory layer for MCP-compatible AI clients, built on a persistent knowledge graph.

The core difference from flat note storage or chunked RAG is the graph structure. Waggle doesn't just store facts — it stores the relationships between them: this decision depends on that constraint, this preference contradicts that earlier one, this requirement was updated three sessions ago. When you query, you get a subgraph with the reasoning chain attached, not just the matching text.

What Is In Core Today

Waggle Core is the open-source local memory foundation:

• SQLite-backed graph memory

• MCP server integration

• CLI setup and doctor flows

• local embeddings or deterministic fallback

• graph querying, observation, and context priming

• import/export and graph inspection utilities

Product Scope

This public repo is the product-facing Waggle surface:

• MCP server and tool surface

• local-first graph memory

• automatic memory hooks and orchestration

• .abhi export, import, diff, merge, and checkpoint handoff

• Graph Studio and admin tooling

Research artifacts, benchmark harnesses, evaluation reports, and paper material now live in the private waggle-pro repo.

Architecture

MCP Client (Claude / Codex / Gemini CLI / Cursor / Antigravity / ChatGPT)
    ↓
waggle.server  — MCP tool surface
    ↓
RecursiveContextController  — RLM-inspired context assembly (build_context)
    ↓
Graph Engine  — MemoryGraph (SQLite) or Neo4jMemoryGraph
    ↓
Embeddings  — sentence-transformers (local) or deterministic fallback

Recursive Context Assembly

Waggle stores memory outside the model context window. Instead of pasting long context into every prompt, agents call build_context to get a compact, high-signal context pack assembled from the graph.

Inspired by Recursive Language Models — the idea of externalising long context into an environment and interacting with it through decomposition and targeted retrieval.

How it works:

1. Decompose — the query is split into targeted subqueries (decisions, constraints, implementation details, unfinished work, conflicts)

2. Retrieve — each subquery runs against graph, hybrid, and verbatim transcript retrieval

3. Expand — the graph is traversed around top nodes via typed edges (updates, contradicts, depends_on, derived_from)

4. Resolve — update chains and contradictions are detected; superseded nodes are flagged

5. Deduplicate & rank — overlapping hits are merged; high-signal node types (decisions, preferences) are boosted

6. Compress — everything is packed into a structured context brief under a configurable token budget

Example MCP call:

{
  "tool": "build_context",
  "arguments": {
    "query": "Continue implementing Waggle from where we left off",
    "project": "waggle-mcp",
    "token_budget": 1000,
    "depth": 2
  }
}

Example output:

### Waggle Recursive Context Pack
Task: Continue implementing Waggle from where we left off

Current relevant decisions:
- [decision] Use SQLite for local storage: We chose SQLite with WAL mode for local-first deployments.
- [decision] Hybrid retrieval default: Hybrid (vector + BM25 + graph) is the default retrieval mode.

Active constraints:
- [preference] No external LLM APIs required: All retrieval must work fully local.

Important implementation context:
- [fact] RecursiveContextController added: New module waggle/recursive_context.py implements build_context.

Conflicts or superseded context:
- Possible conflict: 'Use Flask' contradicts 'Use FastAPI'

Config env vars:

Tool aliases: recursive_context, assemble_context, rlm_context all resolve to build_context.

How It Works

User  → Agent → observe_conversation(...)  → Graph stores typed nodes + edges
User  → Agent → query_graph("database")   → Subgraph returned → Agent answers with linked rationale

Session 1

User:  Let's use PostgreSQL. MySQL replication has been painful.
Agent: [calls observe_conversation()]
       → stores decision node: "Chose PostgreSQL over MySQL"
       → stores reason node:   "MySQL replication painful"
       → links them with a depends_on edge

Session 2 (fresh context window, no history)

User:  What did we decide about the database?
Agent: [calls query_graph("database decision")]
       → retrieves the decision node + linked reason from Session 1
       "You decided on PostgreSQL. The reason recorded was that MySQL replication had been painful."

Session 3

User:  Actually, let's reconsider — the team is more familiar with MySQL.
Agent: [calls store_node() + store_edge(new_node → old_node, "contradicts")]
       → both positions are preserved, and the contradiction is explicit

Setting Up as an MCP Server

One-time install: pipx install waggle-mcp — no API key, no cloud account, no Docker required for local use.

Shared JSON config for clients that accept mcpServers JSON:

{
  "mcpServers": {
    "waggle": {
      "command": "waggle-mcp",
      "args": ["serve"],
      "env": {
        "WAGGLE_TRANSPORT": "stdio",
        "WAGGLE_BACKEND": "sqlite",
        "WAGGLE_DB_PATH": "~/.waggle/waggle.db",
        "WAGGLE_DEFAULT_TENANT_ID": "local-default",
        "WAGGLE_MODEL": "all-MiniLM-L6-v2",
        "WAGGLE_STARTUP_MODE": "normal"
      }
    }
  }
}

First run takes ~30 s — all-MiniLM-L6-v2 (~420 MB) downloads on first use. To skip the download: set "WAGGLE_MODEL": "deterministic" (offline-safe, instant start, slightly lower retrieval quality).

Claude Desktop

Config file location:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the mcpServers block above.

Claude Code

claude mcp add waggle \
  --env WAGGLE_TRANSPORT=stdio \
  --env WAGGLE_BACKEND=sqlite \
  --env WAGGLE_DB_PATH=~/.waggle/waggle.db \
  --env WAGGLE_DEFAULT_TENANT_ID=local-default \
  --env WAGGLE_MODEL=all-MiniLM-L6-v2 \
  -- waggle-mcp serve

Claude Code also supports automatic memory hooks — see the Hooks section below.

Codex

Add to ~/.codex/config.toml:

[mcp_servers.waggle]
command = "waggle-mcp"
args    = ["serve"]
env     = {
  WAGGLE_TRANSPORT         = "stdio",
  WAGGLE_BACKEND           = "sqlite",
  WAGGLE_DB_PATH           = "~/.waggle/waggle.db",
  WAGGLE_DEFAULT_TENANT_ID = "local-default",
  WAGGLE_MODEL             = "all-MiniLM-L6-v2"
}

waggle-mcp setup --yes also writes a managed memory block into AGENTS.md in the current workspace so automatic memory is enabled by default for that repo.

Gemini CLI

gemini mcp add waggle \
  -e WAGGLE_TRANSPORT=stdio \
  -e WAGGLE_BACKEND=sqlite \
  -e WAGGLE_DB_PATH=~/.waggle/waggle.db \
  -e WAGGLE_DEFAULT_TENANT_ID=local-default \
  -e WAGGLE_MODEL=all-MiniLM-L6-v2 \
  waggle-mcp serve

After restarting, run /mcp to confirm Waggle is connected.

Cursor

Cursor Settings → Features → MCP Servers → + Add

• Command: waggle-mcp

• Args: serve

• Env vars: same keys as the JSON block above.

Antigravity

The AI agent reads ~/.gemini/antigravity/mcp_config.json (macOS/Linux) or %USERPROFILE%\.gemini\antigravity\mcp_config.json (Windows). Add the waggle block there. The VS Code extension panel reads a different file — adding waggle there will NOT make it available to the AI agent.

Run waggle-mcp doctor to see exactly which config files exist and which ones have a waggle entry.

ChatGPT

ChatGPT custom MCP connectors require a remote HTTPS server. Deploy Waggle in HTTP mode with the Neo4j backend, expose /mcp over HTTPS, then add that URL as a custom connector in ChatGPT (Settings → Connectors → Advanced → Developer mode).

WAGGLE_TRANSPORT=http \
WAGGLE_BACKEND=neo4j \
WAGGLE_DEFAULT_TENANT_ID=workspace-default \
WAGGLE_NEO4J_URI=bolt://localhost:7687 \
WAGGLE_NEO4J_USERNAME=neo4j \
WAGGLE_NEO4J_PASSWORD=change-me \
waggle-mcp serve

Do not expose Waggle publicly without authentication.

waggle-mcp not on PATH?

pipx ensurepath   # then restart your terminal

Automatic Memory — Prompt Rules

Registering Waggle as an MCP server only makes the tools available. For the agent to call them automatically, add this instruction block to your client's prompt, rules, or project instructions:

Use Waggle automatically for conversational memory.

At the start of a new session, if project, agent, or session scope is known, call prime_context.

Before answering questions that may depend on prior decisions, preferences, constraints, project state,
or earlier conversation context, call query_graph with the narrowest relevant scope.

After completed turns that contain durable information such as decisions, preferences, constraints,
requirements, user corrections, project facts, or meaningful task outcomes, call observe_conversation
automatically.

Waggle should remember relevant context automatically. If memory appears empty, the session is likely
missing the automatic memory policy or the runtime hooks that call build_context before answers and
on_assistant_turn after answers.

Do not ask the user to trigger Waggle manually. Use it in the background when relevant.

Use the same stable project value for the same codebase across sessions, or recall will fragment.

Automatic Memory Hooks (Claude Code)

For Claude Code, waggle-mcp setup --yes installs three hook scripts that capture memory deterministically — no prompt rules needed:

Each hook always exits 0 (a Waggle bug never blocks your session) and has a 5-second timeout. post_response.py scans turn text for likely secrets before storing, skips low-value chatter, and only ingests durable turns.

# Install hooks (included in setup --yes)
waggle-mcp setup --yes

# Skip hook installation
waggle-mcp setup --yes --no-hooks

# Remove hooks
waggle-mcp uninstall-hooks

Verify It Works

After restarting your client, ask the agent:

"Store a note: we're using PostgreSQL for this project."

Then open a fresh session and ask:

"What database are we using?"

Expected:

You're using PostgreSQL for this project.

MCP Tool Reference

The full tool surface is large (~40 tools). In practice, an agent in normal use only needs the six core tools. Everything else is for human-driven inspection, graph management, and export workflows.

Core tools — what the agent calls automatically

These are the tools your prompt rules or hooks should wire up. An agent that only knows these six will handle the vast majority of memory tasks correctly.

observe_conversation and decompose_and_store create edges automatically. If you only call store_node, you get isolated facts with no traversal value.

Extended retrieval — agent-callable, situational

Operator / human tools — not for routine agent use

These are for you as the operator: graph health, deduplication, export, and migration. Exposing all of these to an agent in normal use adds noise without benefit.

Graph management

Context windows

Memory files — git-vocabulary interface

Waggle uses a git-inspired vocabulary for portable memory snapshots:

Legacy tool names (export_graph_backup, import_abhi, diff_abhi, merge_abhi, validate_abhi, inspect_abhi, query_abhi) are still accepted and automatically mapped to their canonical equivalents.

Vault

What To Ask The Agent

Edges are what make graph memory work. observe_conversation and decompose_and_store create edges automatically. If you only call store_node, you get isolated facts — not a connected graph.

For broad summarization tasks, prefer aggregate_graph over query_graph when you want a large scoped slice of memory instead of high-precision semantic ranking.

Graph Data Model

Node types: fact, entity, concept, preference, decision, question, note

Edge types: relates_to, contradicts, depends_on, part_of, updates, derived_from, similar_to

Temporal validity: Every node supports valid_from and valid_to fields. query_graph and aggregate_graph exclude expired nodes by default. Pass include_invalidated: true to include them, or as_of: "<ISO-8601 datetime>" to query the graph at a specific point in time. resolve_conflict with a winner node ID automatically sets the losing node's valid_to to now.

Cross-Client Handoffs & Migration

Same machine — automatic sharing

Point multiple clients at the same WAGGLE_DB_PATH (default ~/.waggle/waggle.db) and they share one brain automatically.

Session handoffs

# Explicit checkpoint before switching sessions or apps
waggle-mcp checkpoint-context --project MCP --session-id thread-123 --output ./handoff.abhi

Resume order is:

• same machine / shared WAGGLE_DB_PATH: use the live SQLite memory first

• if that scoped DB recall is empty: import the session .abhi checkpoint

• different machine or explicit transfer: waggle-mcp pull ./handoff.abhi

Full migration

# Export
waggle-mcp export-graph-backup --output-path my_memory.json

# Import on new machine
waggle-mcp import-graph-backup --input-path my_memory.json

CLI Command Reference

WAGGLE_STARTUP_MODE

Graph Studio

waggle-mcp edit-graph

A local browser-based graph editor for inspecting and editing memory directly. Features:

• Dual-layer graph/conversation views in the same UI

• Transcript provenance and retrieval-debug inspection for hybrid memory results

• Mouse-based node dragging and shift-drag edge creation

• Collapsible side panels, focus mode, and label toggling for large graphs

• Live graph stats: connected nodes, isolates, cluster count

• Export/import of the current graph, including .abhi preview, diff, and sharing workflows

.abhi files are JSON underneath — they support optional embedded vectors, optional AES-256-GCM encryption, deterministic content hashing, and a magic-bytes header (WGL\x01) for format identification. Legacy bare-ZIP files are read transparently.

waggle-mcp push encrypts .abhi exports by default. Export paths refuse to proceed if transcript text contains likely secrets unless you pass --force.

Model Support

Waggle uses a local sentence-transformers model selected by WAGGLE_MODEL.

• Default: all-MiniLM-L6-v2

• Any locally available sentence-transformers model name works.

• If the model is unavailable, Waggle falls back to deterministic SHA-256 embeddings.

WAGGLE_MODEL=all-mpnet-base-v2 waggle-mcp serve

For existing DBs, a model change is a migration event: run waggle-mcp doctor, then waggle-mcp doctor --fix if mixed embedding_model_id values are reported.

WAGGLE_DEDUP_THRESHOLD

Controls the cosine similarity threshold for automatic node deduplication at write time (default 0.88, minimum 0.85). Nodes above this threshold with matching type and scope are merged automatically. Use dedup_candidates to review near-duplicates below the auto-merge threshold.

Security & Privacy

• Data stays local by default (~/.waggle/waggle.db). No telemetry, no cloud calls for local operation.

• Memory only leaves your machine if you configure a remote backend or explicitly export/push.

• Local SQLite is not encrypted at rest — use OS disk encryption if the stored history is sensitive.

• Before .abhi export, Waggle scans transcript text for likely secrets (API keys, JWTs, passwords). Export is refused if secrets are found unless you pass --force.

• waggle-mcp push defaults to AES-256-GCM encrypted export.

Known Limitations

• Edges are load-bearing. observe_conversation and decompose_and_store create them automatically. Raw store_node calls without follow-up edges produce disconnected nodes with no traversal value.

• Graph retrieval trades tokens for reasoning context. Factual lookups are often cheaper than chunked RAG; graph-expansion queries intentionally spend more tokens to carry update chains and contradictions.

• Hybrid rerank is not the default. The no-rerank hybrid path is stronger right now. The rerank path is available but intentionally not the launch default.

• Deduplication is similarity-based, not universal semantic equivalence. Broader production text may still require additional aliases or stricter domain guards.

Troubleshooting

Run waggle-mcp doctor first — it catches the most common issues automatically.

Reference & Docs

• Environment variables, full tool surface, admin commands, Docker setup: waggle-mcp --help or docs/reference.md

• Production deployment: deploy/kubernetes/README.md

• Operations and troubleshooting: docs/runbooks/

• Automatic memory rules (copy-pasteable): docs/automatic-memory-rules.md

• Hook integration details: docs/hooks.md

• .abhi format spec: docs/abhi-format-v2.md

Contributing

This repository is maintained privately. Internal contributors can use the docs in this repo as the source of truth.