For AI Developers

You've rebuilt the same context
47 times this month.

Every new Claude session, every Cursor restart, every LangChain run starts from zero. You paste the same project brief, re-explain the same architecture, re-teach the same preferences. That's not engineering — that's Groundhog Day.

Get a Free API Key See the MCP config ↓

Free tier · No credit card · Works with Claude Desktop, Cursor, Windsurf, LangChain & raw REST

Sound familiar?

A day in the life of an agent with no memory:

session #1 · 9:02 AM

"Here's my stack: Hono on Cloudflare, D1 for storage, Drizzle ORM…" — 1,400 tokens of context you typed yesterday. And the day before.

session #2 · 1:15 PM

Agent suggests the exact fix it already suggested last Tuesday. It solved this bug before. It just doesn't remember solving it.

session #3 · 6:40 PM

Context window fills up mid-task. Everything the agent learned in the last 3 hours silently falls off the end of the buffer.

Context windows are RAM. You're shipping agents with no hard drive.

Bigger context windows don't fix this — they just make the amnesia more expensive. What you need is storage that survives the session: structured, queryable, and shared across every agent you run.

The fix

7 MCP tools. One config block.
Your agent never forgets again.

RJ Knowledge Graph speaks native Model Context Protocol — the same protocol Claude Desktop, Cursor and Windsurf already use. No SDK, no wrapper library, no glue code.

store_memory

Persist a fact, decision or preference. Auto-extracts entities & relationships into the graph — no schema design required.

recall_memory

Hybrid recall: vector similarity plus graph traversal. Finds what keyword search and plain RAG both miss.

explore_graph

Walk the entity graph N hops deep. "What connects Acme Corp to the Q3 pricing decision?" — answered structurally.

ask_knowledge_graph

Full Graph-RAG pipeline: retrieve → rank → synthesize a cited answer with Workers AI. Citations included.

ingest_document

Feed it markdown, docs, notes. Chunking, embedding and entity extraction happen automatically on the edge.

+ namespaces & analytics

Per-agent memory isolation, graph stats, community detection, PageRank centrality — all queryable as tools.

claude_desktop_config.json — the entire integration

{
  "mcpServers": {
    "rj-knowledge-graph": {
      "url": "https://rj-knowledge-graph.pages.dev/mcp",
      "headers": { "X-API-Key": "kg_your_key_here" }
    }
  }
}

That's it. Restart Claude Desktop and your agent has permanent memory. Full quickstart →

Why vector search alone keeps failing you

Embeddings find things that sound similar. Graphs find things that are actually connected. Your agents need both — in one query.

VECTOR-ONLY RAG

Query: "who decided the Q3 pricing?"

Returns 5 chunks that mention pricing. The decision-maker's name lives in a different document that never says "pricing" — so it's invisible. Your agent hallucinates the rest.

HYBRID VECTOR + GRAPH

Same query. Vector search finds the pricing chunks → graph traversal follows DECIDED_BY edges → lands on Dana (CFO), sourced from a doc vector search never touched.

Multi-hop reasoning, backed by citations. No hallucination required.

BGE embeddings via Workers AI

default graph traversal depth

edge latency, 330+ cities

servers for you to run

Machine-readable everything.
Zero lock-in.

Your memory layer shouldn't be a roach motel. Everything that goes in comes out — in the format your tooling expects.

OpenAPI 3.1 spec at /api/openapi.json — codegen clients in any language
llms.txt — your other AI tools can read the docs themselves
Full graph export — JSON, GraphML (Gephi/yEd) or Cypher (Neo4j)
Bitemporal facts — "what did we believe on March 3rd?" is a query, not archaeology

export — take your brain anywhere

# JSON
curl .../api/export/graph
# GraphML for Gephi
curl .../api/export/graph?format=graphml
# Cypher for Neo4j
curl .../api/export/graph?format=cypher

"Yeah, but…" — answered.

Why not just use a bigger context window?

Context windows are working memory — they reset every session and cost tokens on every call. A knowledge graph is long-term storage: write once, recall forever, share across every agent and session. They complement each other; one cannot replace the other.

How is this different from a vector DB + my own glue code?

A vector DB gives you similarity search only. You would still need to build entity extraction, relationship mapping, graph traversal, temporal versioning, namespaces, and an MCP server on top. That is roughly 3–6 weeks of work — this is one config block.

Does it work outside Claude Desktop?

Yes. MCP works with Cursor, Windsurf and any MCP-compatible client. Everything is also exposed as a clean REST API with an OpenAPI 3.1 spec — use it from LangChain, CrewAI, raw Python, or curl.

What about latency in my agent loop?

The whole stack runs on Cloudflare edge (330+ cities). Typical recall round-trip is under 100ms — faster than a single LLM token. Your agent loop will not notice it.

Can I self-verify what is stored?

Yes — the live dashboard shows the full 3D graph, every memory event, and analytics. Plus full export in JSON / GraphML / Cypher at any time. No black box.

You've rebuilt the same context47 times this month.