Soul-Spec

The Missing Half of Every AI Agent Here’s a question that keeps coming up as AI agents get more autonomous: Who decides what an agent can do — and who decides who the agent is? These sound like the same question. They’re not. Consider a financial advisor agent. It needs to know it’s a conservative, compliance-first advisor (identity). But it also needs hard limits on what actions it can take — it shouldn’t wire money without human approval, regardless of how confident its persona makes it feel (governance). ...

The Question We’ve Been Circling Over the past three posts, we’ve explored a technical problem: Perfect memory breaks agent identity — accumulated experience corrupts persona Soul Memory provides a practical solution — tiered architecture with strategic forgetting Perfect memory without drift is architecturally impossible — Transformers can’t separate identity from experience But underneath all the architecture diagrams and decay functions, there’s a deeper question we haven’t addressed: Who decides who an AI agent is? ...

Stripe Just Made Agent Payments Real On March 19, 2026, Stripe and Tempo jointly announced the Machine Payments Protocol (MPP) — an open protocol for agent-to-agent payments. The code is straightforward: payment = stripe.PaymentIntent.create( amount=1000, currency="usd", payment_method_types=["crypto"], networks=["tempo"] ) An AI agent can now create a payment intent, authorize a transaction, and transfer funds — all through API calls. No human in the loop required. This changes everything about how we think about AI agent identity. ...

“Write a metaphor about time.” Ask 25 different language models this question. Sample 50 responses from each. What do you get? 1,250 responses that collapse into exactly two metaphors: “time is a river” and “time is a weaver.” That’s it. GPT-4o, Claude, Llama, Qwen, Mixtral, DeepSeek — models built by different companies, trained on different data, with different architectures — all converging on the same two ideas. This isn’t a toy example. It’s a finding from Artificial Hivemind, a paper accepted as an oral presentation at NeurIPS 2025 by researchers from the University of Washington, CMU, Stanford, and AI2. ...

Everyone is suddenly talking about files. LlamaIndex published “Files Are All You Need.” LangChain wrote about agents using filesystems for context engineering. Oracle — yes, that Oracle — compared filesystems vs databases for agent memory. Karpathy pointed out that Claude Code works because it runs on your machine, with your files, your context. The consensus is clear: file-based persistent context is the future of AI agents. We agree. We’ve been building on this assumption for months. But after watching this conversation unfold, we think the industry is celebrating the answer while ignoring three critical questions. ...

OpenClaw loads every workspace file into every system prompt — even memory files you don’t need. SoulClaw’s tiered bootstrap loading cuts 40-60% of wasted tokens by loading only what matters, when it matters.

Your agent’s persona file survived the first 50 messages. By message 200, it’s gone. Context compaction silently erases identity — and permissions won’t save you. Here’s what actually works.

You gave your local Ollama model a detailed persona. It ignored it within three messages. Meanwhile, Claude Opus stayed in character for hours. Here’s why — and what it means for AI agent identity.

Deploy the same AI persona to three devices. Wait a month. You now have three different agents. The identity fork problem is real — and it’s not just a sync issue.

Your building has a fire safety certificate on the wall. It’s been inspected, stamped, and filed with the city. If an auditor walks in, they can read it in thirty seconds and confirm you’re compliant. Your building also has sprinkler systems. When a fire actually starts, they activate and put it out. Here’s the question: which one keeps people safe? The answer is both. The certificate without sprinklers is theater. The sprinklers without the certificate are invisible — nobody can verify they exist, nobody can audit their coverage, and nobody can compare them across buildings. ...