Agent memory is no longer one feature. It has split into several design camps: raw recall, profile memory, context filesystems, reflective memory, coding-agent memory, and enterprise context APIs. This guide maps the trade-offs, the real architectures, and the hype gap.
Agent-first tools are starting to look like a real software category. The common pattern is simple: products rebuilt around autonomous software users instead of humans. Email, phone numbers, browsers, memory, payments, APIs, and trust layers are all being redesigned around machine operators.
OpenClaw is still one of the most complete personal-agent control planes. But newer systems are getting better in narrower ways. Deep Agents sharpens the harness layer for long-running work. Hermes Agent pushes the persistent self-improving personal-agent thesis harder. OpenViking attacks the deeper context-architecture problem underneath agent memory.
OpenClaw, nanobot, PicoClaw/Clawlet, Agent Zero, ZeroClaw, and memU aren’t one category. This post maps the layers and the real tradeoffs: execution, security posture, packaging, extensibility, and memory economics—plus a comparison matrix and recommendations.