The Vendor Lock-in Problem
Anthropic, OpenAI, and Google are all shipping "memory" features for their AI assistants. Claude has Knowledge Bases (currently rolling out), ChatGPT has Memory, and Gemini has similar capabilities in development.
These features let AI remember context across sessions — your preferences, past conversations, project details. Sounds great, right?
The catch: your data lives in their infrastructure. When you switch from Claude to GPT, or GPT to Gemini, that accumulated knowledge doesn't come with you. You start over.
How Built-in Knowledge Bases Work
Based on leaked system prompts and documentation, here's what we know about Anthropic's Knowledge Bases:
- 01File-based storage — Knowledge Bases are directories of files (likely markdown) that Claude reads and writes internally
- 02Unstructured text — Claude decides what to persist and how to organize it. There's no schema, no consistent field extraction
- 03Token-heavy retrieval — When Claude checks a Knowledge Base, it loads the relevant files into context. More data = more tokens
- 04Single-provider access — Only Claude can read these files. No API, no export, no way to query from GPT or Gemini
The bottom line: Knowledge Bases are convenient but inflexible. Your data becomes part of Claude's ecosystem, not a portable asset you control.
The Endpoints Approach: Structured & Portable
Endpoints takes a different approach. Instead of unstructured text files managed by one AI, you get:
Structured JSON with Schema Enforcement
Every document you scan produces consistent, typed fields. An invoice always has amount, date, vendor. A receipt always has items[], total, tax. No guessing.
REST API — Any Model Can Query
Your endpoints are accessible via standard HTTP. Claude can query them. GPT can query them. Gemini can query them. Your own scripts can query them. The data layer is model-agnostic.
Token-Efficient Retrieval
Instead of loading entire documents into context, your AI queries for specific fields: GET /api/expenses?month=december. Returns only what's needed. Minimal token cost.
Full Data Ownership
Export your data anytime. Switch AI providers without losing your accumulated knowledge. Your endpoints, your control.
The Hybrid Architecture
You don't have to choose between Knowledge Bases and Endpoints. The smartest setup uses both:
Knowledge Base (Index)
- Lightweight summaries
- Endpoint routes & descriptions
- When to query what
Endpoints (Data)
- Full structured records
- Schema-enforced fields
- Queryable via REST API
How it works: Claude checks its Knowledge Base and sees: "User's expense data is at endpoints.work/api/expenses". It makes the API call, gets structured JSON back, and answers your question — without loading hundreds of receipt images into context.
The Knowledge Base stays lightweight (minimal tokens). The heavy data lives in Endpoints (accessible by any AI). Best of both worlds.
Why This Matters Now
The AI landscape is fragmenting. Today you might prefer Claude. Tomorrow, GPT-5 might be better for your use case. Next month, Gemini might offer features you need.
If your data is locked into one provider's Knowledge Base, every switch costs you. You lose context. You re-upload documents. You re-teach preferences.
With a portable data layer, you switch models like changing clothes. Your endpoints persist. Your structured data remains accessible. The new AI picks up where the old one left off.
Start Owning Your Data
Create your first endpoint in minutes. Free tier includes 10GB storage and 50 document scans per month.