Structured Memory for Enterprise AI

A memory palace
for your entire business.

Stop re-explaining your business to AI. Castles gives every model structured, permissioned access to your knowledge — at a fraction of the tokens.

Get Started Free See How It Works
Works with
Claude
Cursor
Windsurf
VS Code
The Problem

Every AI session starts knowing nothing about your business.

You explain your brand voice. Your org structure. Your compliance rules. Then the session ends, and the next one starts from zero. This is context amnesia.

60–80%
of tokens wasted re-explaining context that should already be known
  • Inconsistent outputs across teams using different prompts for the same knowledge
  • No permissions: sensitive docs get stuffed into system prompts with no access control
  • No audit trail: you have no idea what knowledge your AI actually used
  • Scales backwards: the more knowledge you have, the more expensive and fragile it gets
Why Castles

Built for teams that want AI to know their business.

No six-figure enterprise search contract. No data engineering team to maintain a RAG pipeline.

Pasting
into prompts		 Enterprise
search		 Castles
Authored for AI
Native MCP for Claude, Cursor, ChatGPT
Permissioned access
No infra to set up
Stays current automatically
Fits teams of 20–200
Priced for startups
How It Works

Architect your knowledge. Room by room.

Every castle is a structured floor plan of a knowledge domain. Rooms hold the content. A sitemap gives any LLM a map of the entire layout.

Acme Corp ENTERPRISE CASTLE Brand PUBLIC Sales INTERNAL Product INTERNAL Voice Guide Visual Identity Tone Guidelines Pricing 2025 Case Studies Proposals Roadmap Feature Specs Release Notes SITEMAP ~200 tokens · auto-generated TO ANY LLM VIA MCP
01

Build a Castle

Define a knowledge domain: "Product", "Legal", "Brand". The castle is the outer wall on the floor plan. Everything inside it belongs to one structured territory.

02

Divide into Rooms

Each room holds a category of knowledge with its own access tier. "Brand Voice" is public. "Board Minutes" is confidential. Permissions are walls, not suggestions.

03

Fill with Artefacts

The documents, guidelines, and structured data that live inside each room. Add them manually or let AI import and organise them for you.

04

Generate the Sitemap

Castles auto-generates a compact index (~200 tokens) of the entire floor plan. Any LLM reads the map, navigates to the right room, and fetches only what it needs.

Seamless Integration via MCP

One protocol. Every model. Zero disruption.

Castles connects to any LLM through the Model Context Protocol. Your teams keep using the AI tools they already have (Claude, GPT-4, Gemini) and every session silently arrives with the knowledge it needs.

Consistent outputs at scale

50 people using 5 different AI tools all get the same structured knowledge. No more off-brand outputs because someone forgot to paste the style guide into the prompt.

Permissions your legal team will sign off on

Four access tiers enforced at the database level. Compliance docs stay behind the doors they belong behind. No model ever sees knowledge it shouldn't.

Full visibility into what AI accessed

Every knowledge retrieval is logged with who, what, and when. Answer the audit question your board will ask: "What did our AI actually use to generate that?"

No vendor lock-in. No re-engineering.

Standard MCP protocol means you connect once and it works with any model, today and tomorrow. Switch providers without touching your knowledge layer.

Non-technical teams can manage knowledge

The dashboard is built for business users, not engineers. Your brand team updates brand voice. Your legal team manages compliance. No tickets required.

Token costs stay flat as knowledge grows

The sitemap is ~200 tokens regardless of how large your knowledge base gets. Add 10,000 artefacts and the per-session cost doesn't change.

The Experience

Your team won't even notice it's there.

Castles runs in the background. No new tools to learn, no workflow changes. Every AI conversation just starts smarter.

Zero
workflow disruption
147
tokens to deliver full context
3
rooms accessed in a single query
Use Cases

One knowledge layer. Every model.

Castles works wherever your team works. Brand guidelines in Gemini. Sales intel in ChatGPT. Product specs in Claude Code. Onboarding processes in Copilot. Same structured knowledge, zero re-entry.

Marketing rewrites headlines in Gemini — Castles delivers the brand voice guide automatically.

The Product

Your admin controls it. Your whole team uses it.

Structure knowledge, set access tiers, and monitor usage from one dashboard. No code required. Your admin defines the rules, your team gets the context.

Features

Built for enterprise. Loved by engineers.

Every feature serves two audiences: the business leader who needs control and the developer who needs a clean API.

Gatehouse
Tiered Access Control
Four access tiers (Public, Internal, Restricted, Confidential) enforced at the database level via Row Level Security. No application-layer workarounds.
Map Room
Compact Sitemaps
Every castle generates a ~200 token sitemap. The LLM reads the index and fetches only what it needs. Token cost stays flat regardless of knowledge base size.
Drawbridge
Works with Any LLM
Standard MCP server compatible with Claude, GPT-4, Gemini, and open-source models. Plus OpenAI Actions support for ChatGPT Enterprise.
Architect
AI-Native Setup
Describe your knowledge domain and AI scaffolds the castle structure. Import unstructured docs and AI organises them into rooms and artefacts.
Watchtower
Full Audit Trail
Every knowledge retrieval is logged. Know exactly what was accessed, by whom, and when. Enterprise-grade compliance built in from day one.
Archive
Vector Search
Every artefact is embedded. Semantic search across your entire knowledge base with tier-filtered results. Sub-10ms query times at scale.
Enterprise AI Health

Your Enterprise AI isn't broken.
It's suffering from Context Rot.

Every session that starts from scratch compounds the problem: wasted tokens, inconsistent outputs, and knowledge that never sticks. Castles is the cure.

Output Inconsistency
Every team prompts differently, so the same question returns different answers. Castles gives every model one canonical source — consistent answers, whoever's asking.
Knowledge Drift
Models confidently cite stale policies and discontinued products because nothing tells them what's changed. Update a room in Castles once and every session reflects it automatically.
Context Collapse
Contradictory context accumulates until hallucination rates climb and trust erodes. Castles delivers structured rooms on demand — models only see what's accurate and relevant.
Pricing

Simple, predictable pricing.

Start small, scale when you need to. No surprises.

Free
Try Castles with a single workspace.
$0 forever
No credit card required
  • 1 Castle
  • 3 Rooms per Castle
  • 10 Artefacts total
  • 1,000 MCP calls / month
  • OAuth one-click connect to Claude
Get started free
Most Popular
Starter
For teams building real agent memory.
$49/mo
Per organisation
  • 5 Castles
  • Unlimited Rooms
  • Unlimited Artefacts
  • 10,000 MCP calls / month
  • 4-tier access control
  • OAuth + static API keys
  • Email support
Start with Starter
Enterprise
For organisations with custom needs.
Custom
Tailored to your needs
  • Everything in Starter
  • Unlimited Castles
  • Unlimited MCP calls
  • 4-tier access control
  • Custom seat counts
  • Dedicated support
Contact Sales
FAQ

Frequently asked questions

What is Castles?
Castles is a structured knowledge layer for enterprise AI. You organise your business knowledge into Castles (domains), Rooms (categories), and Artefacts (individual documents). Any AI model your team already uses can then access exactly the knowledge it needs, with permissions and an audit trail.
How is this different from RAG or vector search?
RAG systems dump documents into a vector database and retrieve by similarity. That works for simple lookups, but falls apart when you need structure, permissions, or consistency across teams. Castles gives you an organised, permissioned knowledge architecture that any model can navigate via a lightweight sitemap — not a black-box similarity search over a blob of files.
Do I need to change the AI tools my team uses?
No. Castles connects through the Model Context Protocol (MCP), which works with Claude, Cursor, Windsurf, VS Code + Copilot, and any other MCP-compatible client. Your team keeps using the tools they already know. Castles runs in the background.
Is my data secure?
Yes. Every query is scoped by organisation with Row Level Security at the database level. Access tiers (Public, Internal, Restricted, Confidential) control what each API key can see. Every knowledge retrieval is logged with a full audit trail. Your data is never used to train AI models.
What is MCP and do I need to understand it?
MCP (Model Context Protocol) is an open standard that lets AI models connect to external tools and data sources. You don’t need to understand how it works — Castles handles the integration. Just add your connection details to your AI client and you’re live.
How long does setup take?
Minutes. Describe your knowledge domain and AI scaffolds the castle structure for you. You can also import existing documents and AI will organise them into rooms and artefacts. Most teams are live and querying within a single session.
What happens when I hit my plan limits?
You’ll see a warning as you approach your limits. If you reach them, you can upgrade instantly from the dashboard — no downtime, no data loss. Your existing castles and connections stay exactly as they are.
Can non-technical team members manage knowledge?
Absolutely. The dashboard is designed for business users, not engineers. Your brand team updates brand voice, your legal team manages compliance docs, your sales team maintains competitive intel. No code, no tickets, no waiting on engineering.
Which AI models does Castles work with?
Any model that supports MCP. Today that includes Claude (Desktop and Code), Cursor, Windsurf, VS Code + GitHub Copilot, and any other MCP-compatible client. ChatGPT support is coming soon. Because Castles uses an open protocol, new models work automatically — no re-integration needed.