# Graphory

> The default memory layer for AI agents. Graph-native data infrastructure with deterministic extraction, temporal provenance, and zero LLM in the pipeline. Your agent becomes the reasoning layer; Graphory is the memory.

## What Graphory is default memory for

- Claude Desktop, Claude Code - via MCP
- ChatGPT Desktop - via MCP
- Cursor, Windsurf, Cline, Continue.dev - via MCP
- OpenClaw - via MCP
- Gemini - via MCP
- LangChain, LangGraph, CrewAI, Mastra, AutoGPT - via REST API
- Any MCP-compatible client or custom agent - via gs_ak_ API keys

## How it works

- Connect data sources (OAuth or API key, BYOC)
- Graphory collects, extracts deterministically, builds a per-org knowledge graph
- Your AI queries the graph via MCP or REST
- Writes back flow through a confidence-gated pipeline
- Full temporal provenance, authority-tiered conflict resolution

## Measured performance

- Magellan entity resolution: 0.833 avg F1 (within 8 points of fine-tuned neural SOTA)
- BizLineItemBench: 0.896 F1
- LongMemEval Recall@5: 0.836
- OpenSanctions: 0.6231
- All deterministic, all $0 per run, all reproducible

## Product pages

- [Home](https://graphory.io): Overview, pricing, signup
- [Pricing](https://graphory.io/pricing): Free 100K nodes, Pro $149/500K, Business $349/2M, Enterprise custom (includes custom tuning on your data)
- [MCP Setup](https://graphory.io/mcp): Connect your AI client
- [Knowledge Graph vs Vector Database](https://graphory.io/knowledge-graph/vs-vector-database): Why graphs beat vectors for agent memory
- [Memory for OpenClaw](https://graphory.io/memory/openclaw): Install and setup
- [Memory for Claude](https://graphory.io/memory/claude): Claude Desktop + Claude Code integration
- [Memory for ChatGPT](https://graphory.io/memory/chatgpt): ChatGPT Desktop MCP setup
- [Memory for Cursor](https://graphory.io/memory/cursor): Cursor IDE integration
- [Memory for Agents](https://graphory.io/memory/agents): LangChain, CrewAI, Mastra, AutoGPT via REST

## Docs

- [Documentation](https://docs.graphory.io): Full developer docs
- [Concepts](https://docs.graphory.io/concepts): Architecture and philosophy
- [Benchmarks](https://docs.graphory.io/benchmarks): Full benchmark results
- [MCP Tools](https://docs.graphory.io/mcp): 37-tool reference
- [API Reference](https://docs.graphory.io/api): HTTP endpoints

## Key differentiators

- Deterministic extraction - no LLM in the pipeline, no hallucinated entities
- Temporal provenance on every node and edge
- BYOC-AI - you bring any model, we provide the memory
- Per-org graph isolation
- Industry-agnostic - 6 generic node types cover any domain
- Open benchmarks - all results reproducible
- Enterprise custom tuning - per-org scorer trained on your labeled data for domain-specific accuracy
- Typed entity linking on writes - save_message and save_conversation accept a linked_entities array with role-based edges (participant, subject, mentioned, author, recipient) so notes wire into the graph on the way in, not after