The coreaudiod process was using 45G of memory. Digging into it, the leak came from a Claude Code hook playing sound effects. Fixed it by uninstalling PeonPing and switching to Ghostty's native terminal sound effects instead.

Hermes has these advantages over Lobster: model errors get output to the console. The gateway has logs you can check, and crucially, I know where to look. The docs are thorough — no Chinese translation, sure, but they're complete; you can find clear documentation for any issue or command. Each agent has its own separate profile config and gateway, so combinations are more flexible, and the backend can be local or remote. It runs on Docker, so it's use-and-go, with automatic directory mounting. It comes with factory settings that are beginner-friendly — you can get it running smoothly without much configuration. Its built-in security mechanisms are just enough to run basic functionality; unlike Lobster, where you can't use it unless you crank permissions to the max, and maxing them out is unsafe. The difference in how much users have to understand just to get started is night and day. Memory operations are fast, and actions show up in the chat. You can clearly see exactly what she has remembered.

Honestly, writing code with Claude Code is already a blast, but there's one pain point that's been nagging me — I have to keep my eyes glued to the terminal at all times.

OpenClaw Browser supports three modes: openclaw-managed (standalone instance), remote (remote CDP), existing-session (takes over the user's Chrome); grant the agent tool permissions via alsoAllow: [browser]

Instead of bolting on retrieval, MSA embeds memory directly into the attention mechanism itself. That gives you end-to-end long-term memory and breaks past the accuracy ceiling of traditional RAG and the compute-cost bottleneck of the context window.

Playwright is a well-packaged high-level tool, while CDP is the browser's low-level "machine language". For scenarios where OpenClaw directly drives the host machine's browser main Profile, connecting to port 9222 via the CDP (DevTools MCP) protocol is the best choice

On macOS, Chrome remote debugging must be enabled manually by ticking the checkbox at chrome://inspect/#remote-debugging, rather than relying on command-line flags

Solving the macOS Chrome Remote Debugging Port 9222 Problem: SingletonLock Deadlock, Chrome Security Restrictions, rsync Sync + Automator One-Click Launch Solution

Harness Engineering = building an "office" for the Agent instead of continuing to polish the "email wording." From "writing better prompts" to "building the environment," so the agent can run stably and reliably in real production environments

Hermes Agent is a white-box meta AI Agent developed by Nous Research, MIT licensed. Core positioning difference: OpenClaw = a local-first Agent orchestration framework, Hermes = a self-evolving personal operator

Use an LLM to build a persistent wiki instead of traditional RAG, so knowledge is compiled at ingestion time rather than re-derived at query time. Three-layer architecture: Raw Sources (immutable) → Wiki (LLM-maintained) → Schema (CLAUDE.md/AGENTS.md)

Inside a container you need headless:true + noSandbox:true + extraArgs to disable the setuid/namespace sandbox — all three are required. A SIGUSR1 hot restart doesn't fully take effect for the headless field; you need a full container restart

31B parameters, Apache 2.0 license, runs on a Mac Mini, AIME math jumps from 20% to 89% — Google's open-source models have finally turned things around this time. But tool calling is still a weak spot

One-click deployment of 144 AI employees. I took a look at its docs, and only now do I sort of get it. The claim that it "integrates many Agents" is really because the whole open-source project is built out of standalone Agents, one after another — and there are already over a hundred of them. Each folder holds a single Agent.

At first listen, this thing sounds a lot like DayFlow, which I used last September. I could give it another try. I stopped using DayFlow back then because it used local compute for video analysis, and the analysis was way too inaccurate. I was using Qwen's 4B vision model, and the analysis was so inaccurate that I dropped it. I just went through the usage manual in the official docs, and its features are exactly identical to DayFlow, which I used last year.