The latest articles on DEV Community by WAYLAND ZHANG (@waylandz). https://dev.to/waylandz https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3965490%2F879f18ec-5ee7-4a93-bb75-131485598495.png https://dev.to/waylandz en WAYLAND ZHANG Wed, 03 Jun 2026 00:41:12 +0000 https://dev.to/waylandz/i-built-a-persistent-memory-graph-for-my-mac-ai-agent-heres-the-architecture-47fd https://dev.to/waylandz/i-built-a-persistent-memory-graph-for-my-mac-ai-agent-heres-the-architecture-47fd <p>I've been working on a Mac-native agent framework for about a year. One of the hardest problems: making the agent actually remember context across sessions in a way that's <strong>useful</strong>, not just "here's your last 10 messages."</p> <p>What I ended up with is a knowledge graph — entities (people, projects, tools, decisions) with typed relations, stored locally, updated automatically as you interact.</p> <p>When you say "draft the Q3 report like last time," it knows:</p> <ul> <li>What "last time" looks like (the template, the tone, who received it)</li> <li>Who the stakeholders are</li> <li>Which files are relevant</li> </ul> <h2> The rough architecture </h2> <ul> <li> <strong>Agent loop</strong> runs locally on macOS (Swift daemon + Go backend)</li> <li> <strong>LLM-agnostic</strong>: Claude, GPT-4o, Gemini, Ollama all work</li> <li> <strong>Memory tier 1</strong>: SQLite graph with embedding-based fuzzy retrieval</li> <li> <strong>Memory tier 2</strong>: Standard <code>MEMORY.md</code> files per project</li> <li> <strong>Memory tier 3</strong>: Nightly-trained personal memory model, updated while you sleep</li> <li> <strong>Tools</strong>: file ops, browser (Playwright), calendar, terminal, screen</li> <li> <strong>IM bridge</strong>: agent can push updates to Slack/LINE/Feishu while you're away</li> </ul> <h2> The part that surprised me most </h2> <p>The memory graph changed how I use the agent more than any LLM upgrade did.</p> <p>When an agent has persistent structured context, you stop re-explaining everything and start delegating actual work. The shift from "AI as search box" to "AI as coworker" isn't about the model — it's about memory architecture.</p> <h2> What broke along the way </h2> <p><strong>Auto-approval for filesystem ops</strong> — lost a config file on day 3. Everything destructive now requires a human confirm step.</p> <p><strong>Trying to be a general assistant from day one</strong> — it works way better when you give it a specific workflow to own first. I started with "write and send my weekly status report."</p> <p><strong>Ignoring the LLM choice</strong> — for agentic tasks with tool chains, Claude Sonnet/Opus vs GPT-4o makes a meaningful difference in error recovery rate.</p> <h2> The project </h2> <p>I turned this into <strong>Kocoro</strong> (open-sourced the core runtime: <a href="https://github.com/Kocoro-lab/ShanClaw" rel="noopener noreferrer">github.com/Kocoro-lab/Kocoro</a>). It's a Mac-native AI agent — local-first, memory graph built in, IM notifications, LLM-agnostic.</p> <p>Just opened a closed beta — drop a comment if you want an invite.</p> <p><strong>What are others using for persistent agent memory on local setups?</strong></p> ai agents showdev llm