HN Showcase
Back to browse
Base Layer – Open-source behavioral compression from any text

Base Layer – Open-source behavioral compression from any text

by agulaya24·Mar 10, 2026·1 point·0 comments
Visit ProjectView on HN

AI Analysis

●●●BangerBig BrainRabbit HoleZero to One

Models behavior instead of facts — genuinely different from standard RAG memory.

Strengths
  • •Every claim traces back to source facts with verification hashes.
  • •Tested on diverse corpora from journals to 350k-word shareholder letters.
  • •Four-step pipeline compresses text into three-layer identity models.
Weaknesses
  • •Still relies on Claude API for extraction — not fully local.
  • •Niche audience — most developers don't need behavioral identity modeling.
Category
AI/ML
Target Audience

AI developers building memory systems or personalized agents

Similar To

Mem0 · Zep · Standard RAG memory systems

Post Description

I went down the rabbit hole of AI memory, and this came out the other end.

Beliefs, behaviors, tensions, and contradictions extracted from conversations, journals, and published text, compressed into an identity brief that any model or memory system can use. An extracted operating guide for AI, where every claim traces back to source facts.

All research, benchmarks, documentation, examples are available on the website and in the github. This has been tested on as little at 8 Personal Journal Entries from a secondary subject, my own gpt conversations exports (30K+ Messages), and on large document corpora like Warren Buffett's Annual Shareholder Letters (350k words), Howard Marks Investment Memos (600K words), and dense autobiographies from Franklin, Douglass, Roosevelt, and Wollstonecraft.

Pipeline currently uses Claude. API costs are <$1 for small data sets and <$5 for large ones, from fact extraction to final brief assembly.

Very interested in feedback, happy to go deeper in the comments on evolution, struggles, research, and future improvements.

Similar Projects

AI/ML●●Solid

Experience-engine – reflection-based memory layer for local LLMs

Turns chat history into structured 'belief' and 'cognitive pattern' blocks you can inject into prompts, with simple APIs like run_reflection and run_synthesis that read like a research prototype. It's smart about separating V1 (domain beliefs) from V2 (transferable cognitive patterns), but it's clearly early-stage — tiny repo, Ollama-only workflow, and few commits mean you should treat it as an experimental MVP rather than a drop-in production memory system.

Big BrainNiche GemShip It
ashishluthara
313mo ago