waggle-mcp 0.1.3

MCP server that gives LLMs persistent graph-structured memory

Persistent, structured memory for AI agents — backed by a real knowledge graph.
Your LLM remembers facts, decisions, and context across every conversation.

---

Why waggle-mcp?

Most LLMs forget everything when the conversation ends.
waggle-mcp fixes that by giving your AI a persistent knowledge graph it can read and write through any MCP-compatible client.

Without waggle-mcp	With waggle-mcp
"What did we decide about the DB schema?" → ❌ no idea	✅ Recalls the decision node, when it was made, and what it contradicts
Context stuffed into a 200k-token prompt	Compact subgraph — only relevant nodes retrieved
Flat bullet-list memory	Typed edges: relates_to, contradicts, depends_on, updates…
One session, one agent	Multi-tenant, multi-session, multi-agent

---

Demo

---

Quick start — 30 seconds

pip install waggle-mcp
waggle-mcp init

The init wizard detects your MCP client, writes its config file, and creates the database directory — no JSON editing required. Supports Claude Desktop, Cursor, Codex, and a generic JSON fallback.

After init, restart your MCP client and your AI has persistent memory.
No cloud service. No API key. Semantic search runs fully locally.

---

How it works

Memory doesn't just get stored — it flows through a lifecycle:

You talk to your AI
        │
        ▼
  observe_conversation()          ← AI drops the turn in; facts are auto-extracted
        │
        ▼
  Graph nodes are created         ← "Chose PostgreSQL" becomes a decision node
  Edges are inferred              ← linked to the "database" entity node
        │
        ▼
  Future conversation starts
        │
        ▼
  query_graph("DB schema")        ← semantic search finds the node from 3 sessions ago
        │
        ▼
  AI answers with full context    ← "You decided on PostgreSQL on Apr 10, here's why…"

Every node carries semantic embeddings computed locally using all-MiniLM-L6-v2 — a fast, lightweight model that runs entirely on-device with no API key or network call required. This means semantic search works offline, costs nothing per query, and keeps your data private.

---

See it in action

Here's a concrete before/after for a developer using the AI daily:

Session 1 — April 10

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 depends_on edge

Session 2 — April 12 (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 April 10

       "You decided on PostgreSQL on April 10. The reason recorded was
        that MySQL replication had been painful."

Session 3 — April 14

User:  Actually, let's reconsider — the team is more familiar with MySQL.
Agent: [calls store_node() + store_edge(new_node → old_node, "contradicts")]
       → conflict is flagged automatically; both positions are preserved in the graph

The agent never needed explicit instructions to remember or retrieve — it called the right tools based on the conversation, and the graph gave it the right context.

---

The magic tool: observe_conversation

This is the tool you'll use most. You don't have to manually store facts — just tell the agent to observe each conversation turn and it handles the rest.

observe_conversation(user_message, assistant_response)

Under the hood, it:

1. Extracts atomic facts from both sides of the conversation
2. Deduplicates against existing nodes using semantic similarity
3. Creates typed edges between related concepts
4. Flags contradictions with existing stored beliefs

No instructions needed. No schema to define. Just observe.

---

Temporal queries — a solved problem most memory systems skip

Most memory systems answer "what do you know about X?" — but can't answer when you learned it or how knowledge changed over time.

waggle-mcp understands temporal natural language natively:

Query	What happens
query_graph("what did we decide recently")	Filters nodes updated in the last 24–48h
query_graph("what was the original plan")	Retrieves the earliest version of relevant nodes
query_graph("what changed last week")	Returns a diff of nodes created/updated in that window
graph_diff(since="48h")	Explicit changelog: added nodes, updated nodes, new conflicts

This is built on timestamped nodes + temporal phrase parsing — no vector-clock complexity, but enough to reconstruct a meaningful timeline of decisions.

---

Memory model

Node types — what gets stored:

Type	Example
fact	"The API uses JWT tokens"
preference	"User prefers dark mode"
decision	"Chose PostgreSQL over MySQL"
entity	"Project: waggle-mcp"
concept	"Rate limiting"
question	"Should we add GraphQL?"
note	"TODO: add integration tests"

Edge types — how nodes connect:

relates_to · contradicts · depends_on · part_of · updates · derived_from · similar_to

---

MCP tools

Your AI calls these directly — you don't need to use them manually.

Tool	What it does
observe_conversation	Drop a conversation turn in — facts auto-extracted, stored, and linked
query_graph	Semantic + temporal search across the graph
store_node	Manually save a fact, preference, decision, or note
store_edge	Link two nodes with a typed relationship
get_related	Traverse edges from a specific node
update_node	Update content or tags on an existing node
delete_node	Remove a node and all its edges
decompose_and_store	Break long content into atomic nodes automatically
graph_diff	See what changed in the last N hours
prime_context	Generate a compact brief for a new conversation
get_topics	Detect topic clusters via community detection
get_stats	Node/edge counts and most-connected nodes
export_graph_html	Interactive browser visualization
export_graph_backup	Portable JSON backup
import_graph_backup	Restore from a JSON backup

---

Installation

Local / development (SQLite, no extra services)

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
waggle-mcp init        # ← writes your client config automatically

Three key variables for local mode:

Variable	What it does
WAGGLE_BACKEND=sqlite	Local file DB, zero setup
WAGGLE_TRANSPORT=stdio	Connects to desktop MCP clients
WAGGLE_DB_PATH	Where the graph is stored (default: memory.db)

Production (Neo4j backend)

pip install -e ".[dev,neo4j]"

Then run the server:

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

Docker

docker build -t waggle-mcp:latest .

docker run --rm -p 8080:8080 \
  -e WAGGLE_TRANSPORT=http \
  -e WAGGLE_BACKEND=neo4j \
  -e WAGGLE_DEFAULT_TENANT_ID=workspace-default \
  -e WAGGLE_NEO4J_URI=bolt://host.docker.internal:7687 \
  -e WAGGLE_NEO4J_USERNAME=neo4j \
  -e WAGGLE_NEO4J_PASSWORD=change-me \
  waggle-mcp:latest

---

Manual client configuration

If you prefer to edit config files directly, or init doesn't cover your client:

Claude Desktop — claude_desktop_config.json

{
  "mcpServers": {
    "waggle": {
      "command": "/path/to/.venv/bin/python",
      "args": ["-m", "waggle.server"],
      "env": {
        "PYTHONPATH": "/path/to/waggle-mcp/src",
        "WAGGLE_TRANSPORT": "stdio",
        "WAGGLE_BACKEND": "sqlite",
        "WAGGLE_DB_PATH": "~/.waggle/memory.db",
        "WAGGLE_DEFAULT_TENANT_ID": "local-default",
        "WAGGLE_MODEL": "all-MiniLM-L6-v2"
      }
    }
  }
}

Codex — codex_config.toml

[mcp_servers.waggle]
command = "/path/to/.venv/bin/python"
args    = ["-m", "waggle.server"]
cwd     = "/path/to/waggle-mcp"
env     = {
  PYTHONPATH                     = "/path/to/waggle-mcp/src",
  WAGGLE_TRANSPORT         = "stdio",
  WAGGLE_BACKEND           = "sqlite",
  WAGGLE_DB_PATH           = "~/.waggle/memory.db",
  WAGGLE_DEFAULT_TENANT_ID = "local-default",
  WAGGLE_MODEL             = "all-MiniLM-L6-v2"
}

A pre-filled example is in codex_config.example.toml.

---

Environment variables

Click to expand full reference

Core

Variable	Default	Description
WAGGLE_BACKEND	sqlite	sqlite or neo4j
WAGGLE_TRANSPORT	stdio	stdio or http
WAGGLE_MODEL	all-MiniLM-L6-v2	sentence-transformers model (local inference)
WAGGLE_DEFAULT_TENANT_ID	local-default	default tenant
WAGGLE_EXPORT_DIR	—	optional export directory

SQLite

Variable	Default	Description
WAGGLE_DB_PATH	memory.db	path to the SQLite file

HTTP service

Variable	Default	Description
WAGGLE_HTTP_HOST	0.0.0.0	bind host
WAGGLE_HTTP_PORT	8080	bind port
WAGGLE_LOG_LEVEL	INFO	log level
WAGGLE_RATE_LIMIT_RPM	120	global rate limit (req/min)
WAGGLE_WRITE_RATE_LIMIT_RPM	60	write-tool rate limit
WAGGLE_MAX_CONCURRENT_REQUESTS	8	concurrency cap
WAGGLE_MAX_PAYLOAD_BYTES	1048576	max request size
WAGGLE_REQUEST_TIMEOUT_SECONDS	30	per-request timeout

Neo4j

Variable	Description
WAGGLE_NEO4J_URI	Bolt URI, e.g. bolt://localhost:7687
WAGGLE_NEO4J_USERNAME	Neo4j username
WAGGLE_NEO4J_PASSWORD	Neo4j password
WAGGLE_NEO4J_DATABASE	Neo4j database name

---

Admin commands

# Create a tenant
waggle-mcp create-tenant --tenant-id workspace-a --name "Workspace A"

# Issue an API key (raw key returned once — store it securely)
waggle-mcp create-api-key --tenant-id workspace-a --name "ci-agent"

# List keys for a tenant
waggle-mcp list-api-keys --tenant-id workspace-a

# Revoke a key
waggle-mcp revoke-api-key --api-key-id <id>

# Migrate SQLite data → Neo4j
WAGGLE_BACKEND=neo4j WAGGLE_NEO4J_URI=bolt://localhost:7687 \
WAGGLE_NEO4J_USERNAME=neo4j WAGGLE_NEO4J_PASSWORD=change-me \
  waggle-mcp migrate-sqlite --db-path ./memory.db --tenant-id workspace-a

---

Kubernetes & observability

Full production deployment assets are in deploy/:

Path	What's inside
deploy/kubernetes/	Deployment, Service, Ingress (TLS), NetworkPolicy, HPA, PDB, cert-manager, ExternalSecrets — see deploy/kubernetes/README.md
deploy/observability/	Prometheus scrape config, Grafana dashboard, one-command Docker Compose observability stack

---

Runbooks

Operational runbooks are in docs/runbooks/:

• API key rotation — zero-downtime create-then-revoke
• Incident response — Neo4j down, OOM, rate storm, auth failures
• Backup & restore — manual and automated drill
• Tenant onboarding — new tenant checklist
• Secret management — External Secrets + cert-manager

---

Testing

.venv/bin/pytest -q

Coverage: graph CRUD, deduplication, conflict detection, tenant isolation, backup/import, stdio MCP, HTTP auth/health/metrics, payload limits.

# End-to-end backup/restore drill
WAGGLE_HOST=http://localhost:8080 WAGGLE_API_KEY=<key> \
  ./scripts/backup_restore_drill.sh

# Load test (p50/p95/p99 latency report)
WAGGLE_API_KEY=<key> ./scripts/load_test.sh --medium

---

Architecture

waggle-mcp
├── Core domain    graph CRUD · dedup · local embeddings · conflict detection · export/import
├── Transport      stdio MCP (Codex/Desktop) · streamable HTTP MCP (Kubernetes)
└── Platform       config · auth · tenant isolation · rate limiting · logging · metrics

Backend:

• Local/dev → SQLite (zero config, instant start)
• Production → Neo4j (WAGGLE_TRANSPORT=http requires WAGGLE_BACKEND=neo4j)

---

Project layout

waggle-mcp/
├── assets/                   ← banner + demo SVG
├── deploy/
│   ├── kubernetes/           ← full K8s manifests + guide
│   └── observability/        ← Prometheus + Grafana stack
├── docs/runbooks/            ← operational runbooks
├── scripts/
│   ├── load_test.py / .sh
│   └── backup_restore_drill.py / .sh
├── src/waggle/         ← server, graph, neo4j_graph, auth, config …
├── tests/
├── Dockerfile
├── pyproject.toml
└── README.md

---

License

MIT — see LICENSE.