Four hooks.
Crash-resistant
by design. 四个 hook，
从设计上
抗崩溃。

Each Claude Code project has its own auto-memory directory at ~/.claude/projects/<sanitized-cwd>/memory/. Hooks resolve this path from $CLAUDE_PROJECT_DIR at runtime, so the same Claude Code instance running in different projects keeps separate checkpoints — never confused, never overwriting each other. 每个 Claude Code 项目有自己的 auto-memory 目录 ~/.claude/projects/<sanitized-cwd>/memory/。hook 运行时从 $CLAUDE_PROJECT_DIR 解析路径——同一个 Claude Code 实例在不同项目下跑，checkpoint 各自独立，互不混淆。

# What a project's memory directory looks like
~/.claude/projects/<project>/memory/
├── user_role.md                            ← existing auto-memory
├── feedback_doc_style.md                   ← existing auto-memory
├── project_sdk_status.md                   ← existing auto-memory
├── reference_build_commands.md             ← existing auto-memory
├── project_session_20260501_status.md      ← Claude writes (semantic)
└── project_session_20260501_autosnap.md    ← Stop hook writes (machine)

Two files, one purpose split into two layers. Both sit next to your existing memory entries — auto-memory reads them like any other project memory. Your user_*.md, feedback_*.md, project_*.md, reference_*.md are all left alone. 两个文件，一种用途拆成两层。都跟你现有的 memory 条目放在同一个目录里——auto-memory 读它们跟读其他 project memory 一样。你的 user_*.md / feedback_*.md / project_*.md / reference_*.md 一概不动。

_status.md is the curated narrative — task list, in-progress action, next steps, key context. Refreshed when you talk to Claude and the 30-min reminder fires. _autosnap.md is the objective fallback — written by a shell script with whatever git, jq, and tail can see. Refreshed at most once every 60 minutes by the Stop hook, regardless of whether you are talking. Recovery prefers _status.md; falls back to _autosnap.md when the status file is stale or missing. _status.md 是精炼的叙述——任务列表、进行中、下一步、关键上下文。user 说话且 30 分钟 reminder 触发时刷新。_autosnap.md 是客观兜底——shell 脚本拿 git / jq / tail 能看到的全部信息写下来。Stop hook 最多每 60 分钟重写一次，跟 user 说不说话无关。崩溃恢复时先看 _status.md；它过期或丢了再看 _autosnap.md。

How a new session picks up where the last one left off 新 session 怎么接上上次的进度

Open Claude Code (or resume a paused session, run /clear, or auto-resume after compact). The SessionStart hook fires before the model sees your first message. It injects a system-reminder telling Claude where today's checkpoint is and where the most recent prior-day checkpoint is, so Claude reads them on its own — no user action needed. 打开 Claude Code（或恢复暂停的 session、跑 /clear、压缩后自动恢复），SessionStart hook 在模型看到你第一条消息之前自动触发。它注入 system-reminder 告诉 Claude 今天 checkpoint 在哪、最近一份历史 checkpoint 在哪——Claude 自己会去读，你什么都不用做。

In the same project directory, prior context is automatically re-established. The 30-minute reminder hook then keeps the file fresh as work continues. 在同一个项目目录下，上次的上下文自动接续。然后 30 分钟提醒 hook 让文件随着工作进展持续刷新。