🧠 MollyGraph

Local-first graph memory for AI agents.

Quick Start · How It Works · MCP · API · Config

---

Most agent memory is a vector store with a wrapper. Same extraction quality forever. No structure. No relationships. No local-first story.

MollyGraph is a local graph + vector memory core for AI agents. Ingest text -> extract entities and relationships with GLiNER2 -> store them locally in Ladybug -> serve them through MCP, HTTP, and the SDK.

---

What Makes This Different

• Local-first memory core — Embedded Ladybug graph storage plus local vector storage by default. No mandatory database daemon for the main path.
• GLiNER2-first extraction — The default product path is built around local structured extraction, not cloud reasoning.
• Speaker-anchored ingestion — Each message is processed individually with the speaker as anchor entity, which is the main defense against bad chat graphs.
• Parallel retrieval — Graph exact-match and vector similarity run together, then merge into one response.
• Simple local embeddings — Default embedder is Snowflake/snowflake-arctic-embed-s, with nomic-embed-text via Ollama as the optional local alternative.
• Experimental features still available — audit chains, GLiREL, training loops, and decision traces are being kept behind explicit flags instead of defining the core product.

---

Quick Start

git clone https://github.com/brianmeyer/mollygraph.git
cd mollygraph
./scripts/install.sh
./scripts/start.sh

API at http://127.0.0.1:7422. Auth: Bearer dev-key-change-in-production.

./scripts/install.sh creates service/.env from service/.env.example if it does not exist yet, and installs the runtime into service/.venv.

Default local stack:

• MOLLYGRAPH_GRAPH_BACKEND=ladybug
• MOLLYGRAPH_VECTOR_BACKEND=ladybug
• MOLLYGRAPH_EMBEDDING_ST_MODEL=Snowflake/snowflake-arctic-embed-s

Neo4j is still available for legacy and experimental workflows, but it is no longer the intended default.

---

How It Works

  ingest text
       │
       ▼
  Speaker-anchored extraction (GLiNER2)
  Per-source confidence gates
       │
       ├──▶ Ladybug graph
       ├──▶ Ladybug vector index
       └──▶ MCP / HTTP / SDK query surface

Default query path: graph exact + vector similarity fire in parallel -> merge -> serve.

Experimental path: audit, training, decision traces, and extra enrichment layers are still present, but they are not part of the default stripped-down runtime.

---

MCP Integration

Works with Claude, OpenClaw, Cursor, Ollama-adjacent local agents, or any MCP client.

{
  "mcpServers": {
    "mollygraph": {
      "command": "mollygraph-mcp",
      "args": ["--base-url", "http://localhost:7422", "--api-key", "YOUR_KEY"]
    }
  }
}

Core tools: add_episode · search_facts · search_nodes · get_entity_context · delete_entity · prune_entities · get_queue_status

Legacy tools like audit and training remain available only when the active backend supports them.

---

Python SDK

pip install "git+https://github.com/brianmeyer/mollygraph.git#subdirectory=sdk"

from mollygraph_sdk import MollyGraphClient

client = MollyGraphClient(base_url="http://localhost:7422", api_key="YOUR_KEY")
client.ingest("Sarah joined the ML team at Acme Corp.", source="slack", speaker="Brian")
result = client.query("What do we know about Sarah?")

---

HTTP API

Endpoint	Method	Description
/health	GET	Service health
/ingest	POST	Ingest text (supports speaker field)
/query	GET	Parallel graph+vector search
/entity/{name}	GET	Entity context (2-hop neighborhood)
/entity/{name}	DELETE	Delete entity + relationships
/entities/prune	POST	Bulk prune + orphan detection
/metrics/dashboard	GET	Unified health JSON
/stats	GET	Graph/vector/runtime stats

Neo4j-oriented endpoints like audit, training, and the legacy maintenance cycle remain in the codebase but are intentionally gated when the Ladybug backend is active.

---

Configuration

100% env-var driven. See service/.env.example for all options. Key settings:

# Local-first defaults
MOLLYGRAPH_GRAPH_BACKEND=ladybug
MOLLYGRAPH_VECTOR_BACKEND=ladybug
MOLLYGRAPH_EMBEDDING_ST_MODEL=Snowflake/snowflake-arctic-embed-s

# Optional local alternate
MOLLYGRAPH_EMBEDDING_OLLAMA_MODEL=nomic-embed-text

# Extraction confidence (per source)
MOLLYGRAPH_EXTRACTION_CONFIDENCE_SESSION=0.55
MOLLYGRAPH_EXTRACTION_CONFIDENCE_EMAIL=0.45
MOLLYGRAPH_EXTRACTION_CONFIDENCE_DEFAULT=0.4

# Neo4j remains available for legacy/full workflows
# NEO4J_URI=bolt://localhost:7687

---

Roadmap

The authoritative roadmap lives in service/BACKLOG.md. service/DECISION_TRACES_PLAN.md is a later-phase experimental plan, not part of the default product path.

Top priorities right now:

• Finish the Ladybug local core for the default ingest/query/vector path
• Keep the default runtime small and trustworthy for non-developers
• Harden MCP, HTTP, and SDK behavior around the simplified memory core
• Add source-routed extraction where it improves graph quality without re-complexifying the product
• Keep advanced audit/training/decision-trace work clearly optional

Current feature status:

• Ladybug graph backend — active default
• Ladybug vector backend — active
• GLiNER2 local extraction — active
• Snowflake local embeddings — active default
• Audit, training, and decision surfaces — legacy/experimental, being gated behind backend capability

---

License

MIT

---

Structured local memory for agents.