HAL

A personal knowledge system with semantic search. Tags: MCP, RAG, Data Sovereignty.

The Problem

Every conversation with ChatGPT or Claude is an investment, but it is in their moat, not mine. Each interaction builds context, and more context locks you in.

With features like memory, project instructions, and project files, your dependence becomes increasingly entrenched with your provider of choice. Those are valuable capabilities, but switching costs mount over time.

Context is infrastructure, not just a repository of files. LLMs compel you to document tacit knowledge because you want better answers, so you keep feeding them context that may have only lived in your head before.

How I Built It

I built HAL to prove that context sovereignty is technically feasible. The substrate, the cumulated knowledge and processes that expound it, lives with me.

Storage: Local Markdown files with Time Machine backups.
Search: Weaviate for semantic and keyword search.
Embeddings: Mistral.
Interface: An MCP server that any compatible AI can connect to.

What I Am Learning

HAL works as a weekend build and demonstrates that there is a there there. The question is not whether LLMs are useful; they clearly are. I want to make sure I am not locked into any single one of them.

If this resonates, keep important notes in local files where you can own them, copy valuable AI conversations to your own documents, and use multiple AI platforms rather than going all-in on one.

Roadmap

Local stack: Migrate to fully local and open-source Weaviate instance and embeddings.
Models: Test open-source LLMs.
Capture: Maybe add voice-note transcription.