Soul-Spec

Anthropic just published the largest qualitative AI study ever conducted. 80,508 people. 159 countries. 70 languages. One week. And the results flip the dominant narrative about what AI users actually care about. The headline finding is deceptively simple: people don’t want AI that does more. They want AI they can trust. The Study The “81k Interviews” project used Claude-based AI interviewers to conduct structured conversations with participants worldwide. Each interview adapted its follow-up questions based on responses — a hybrid approach that captures both the scale of surveys and the depth of qualitative research. ...

On March 31, 2026, security researcher Chaofan Shou discovered something Anthropic probably didn’t want the world to see: the entire source code of Claude Code — Anthropic’s official AI coding CLI — sitting in plain sight on the npm registry via a .map file bundled into the published package. The model wasn’t leaked. The weights are safe. But everything else — the agent architecture, the multi-agent orchestration, the memory system, the internal feature flags — all of it was exposed. ...

The AI industry loves naming eras. We had the prompt engineering era. Then came context engineering. Now we’re entering what may be the most consequential shift yet: harness engineering. Each stage represents a fundamental change in what we’re designing when we build AI systems. And each stage demands a different kind of specification. Stage 1: Prompt Engineering — Talking to the Model The first era was about learning to talk to AI. We crafted system prompts, experimented with role-playing instructions, and discovered that saying “think step by step” actually worked. ...

TL;DR We launched ClawSouls Registry — an open registry where anyone can submit AI agent personas via GitHub Pull Request. Every submission is automatically verified by SoulScan (53 safety patterns) before it can be merged. No other AI agent registry does automated safety verification. That’s our differentiator. The Problem AI agents are everywhere — Claude Code, Cursor, Windsurf, Copilot. Each one can be personalized with system prompts, personality files, or persona definitions. But there’s no safe, standardized way to share these personas. ...

AI agents forget who they are every session. Soul-Driven AI fixes that — persistent identity, verified safety, community-built personas. Here’s what it means and why it matters.

We tested three persona-level safety approaches on LLMs with safety training surgically removed. Rules alone: 28%. Governance alone: 44-61%. Combined identity + governance: 100%. Here’s what we learned.

We tested whether structured persona files can restore safety in abliterated LLMs — models where safety guardrails have been surgically removed. The results reveal a striking asymmetry that challenges conventional thinking about AI safety.

New Paper: The Forgetting Problem We’ve published a new preprint exploring a counterintuitive idea: the better an AI agent’s memory, the worse its identity becomes. 📄 Read the paper on Zenodo (CC-BY 4.0, open access) The Memory-Identity Paradox Every major AI agent framework is racing to build better memory. MemGPT, Mem0, A-Mem, MemoryBank — all optimize for remembering more, longer, more accurately. But we identified a fundamental tension: The more faithfully an agent remembers its experiences, the more vulnerable its intended identity becomes to experiential contamination. ...

Your AI Agent Remembers Everything. That’s the Problem. Every agent framework is racing to build better memory. MemGPT, Mem0, A-Mem — they all want your agent to remember more, longer, better. But here’s a question nobody’s asking: what happens to your agent’s personality when it remembers too much? Humans Forget for a Reason In psychology, there’s a concept called adaptive forgetting. Your brain doesn’t just lose information by accident — it actively suppresses memories that would interfere with your ability to function. ...

The Problem: Your Agent Either Remembers Everything or Nothing Every AI agent developer faces the same dilemma: No memory → Your agent forgets everything between sessions. Every conversation starts from zero. Full memory → Your agent remembers everything with perfect fidelity. Including that one time a user was hostile. Including outdated decisions. Including noise from 6 months ago that drowns out yesterday’s critical update. Neither is right. Humans solved this millions of years ago: we remember what matters and forget what doesn’t. Not perfectly — but well enough to maintain a coherent identity while adapting to new experiences. ...