Kocoro 是一个引擎级的开源 Mac AI agent 框架。它通过本地 agent 读取用户的过往会话、文件、应用、浏览器、屏幕和终端,将有用事实压缩为记忆,使 agent 无需重复上下文即可继续工作。安全模型以本地优先:工具操作需授权,危险命令被拦截或二次确认,行为可审计,秘密自动脱敏,记忆/同步默认不上传。其“情景记忆”将历史会话转化为项目事实、决策、协作者、截止日期和习惯。每晚自动将工作日蒸馏成本地知识图谱,次日直接恢复进度。
A longer context window does not solve the real memory problem in AI work.
Kocoro just made AI memory a local Mac feature.
It's an open-source Mac AI agent framework at the engine level.
Kocoro works by running a local agent on your Mac that can read your past sessions, files, apps, browser, screen, and terminal, then compress useful facts into memory so it can continue work without you repeating everything.
Its security model is mostly local-first control: tool actions need permission, risky commands are blocked or re-asked, actions are audit-logged, secrets are auto-redacted, and memory/session sync is opt-in rather than always uploaded.
Its Episodic Memory turns past sessions into selected project facts, decisions, collaborators, deadlines, and habits, so the agent can resume work like a teammate rather than a help desk ticket.
Every night it distills your workday into a local knowledge graph - projects, decisions, open tasks. Next morning it picks up exactly where you left off. No context re-pasting.
Github links in comments