DEV Community

Cover image for Keep every Claude Code chat forever, at zero token cost
Amandeep Singh
Amandeep Singh Subscriber

Posted on Edited on AI-assisted

Keep every Claude Code chat forever, at zero token cost

Claude Code deletes CLI and VS Code transcripts after 30 days by default (cleanupPeriodDays).
It also files them under your folder path as raw JSON, which is not something you want to grep when you are looking for "that thing we decided three weeks ago".

I wanted a permanent, readable copy of every chat, so I built a plugin: session-archive.

What it does

One folder per chat:

<archive>/<project>/<YYYY-MM-DD_HHMM>_<id>/
  chat.md        the conversation as you saw it, with tool calls and images inline
  meta.json      title, branch, model, dates, counts
  raw/           byte-exact transcript, subagents, tool results
  artifacts/     every version of every file Claude wrote or edited
  attachments/   every image shown in the chat
Enter fullscreen mode Exit fullscreen mode

a chat.md open in my editor, showing a tool call with its collapsed output and an inline image

It runs from hooks, so it costs zero model tokens.
It works in the CLI, the VS Code extension and the Code tab of the Claude desktop app.

Install

claude plugin marketplace add singhAmandeep007/claude-session-archive
claude plugin install session-archive@claude-session-archive --config archive_dir=~/claude-archive
Enter fullscreen mode Exit fullscreen mode

Then start a new session. That is it.
Point archive_dir at any folder you like. A cloud-synced folder works too, which gives you a backup for free.

You need Python 3.9+ as python3 and git on your PATH. Nothing to pip install.

Two commands

  • /session-archive:save saves the current chat now. /session-archive:save all archives every chat still on disk.
  • /session-archive:import restores archived chats byte-exact into Claude Code, so claude --resume works on a chat that was already deleted. It never overwrites a live transcript that differs.

session-archive:save showing the one-line report in the terminal or desktop app

The part I got wrong at first

My first version answered those slash commands from a hook.
Hooks can only answer a command by blocking it, and Claude Code always shows a blocked prompt as a warning: "A hook blocked your prompt".
It worked perfectly and looked like an error every single time.

the old "A hook blocked your prompt" banner, from my earlier screenshot

Version 1.6.0 fixes that with a Claude Code mod (hooks/register.ts, Claude Code 2.1.287 or newer).
A mod can answer a command directly, so the report shows up as normal command output, still without a model turn.
Older Claude Code versions keep using the hook path.

Small things that matter

  • Saves after a reply run at most every 10 minutes per session (min_sync_minutes), because cloud-synced folders re-upload whole files and transcripts get big.
  • Each session only ever gets one folder, even if you start it outside a repo and move into one later.
  • A damaged transcript (bad bytes, record shapes it does not know) is skipped record by record instead of stopping the archive.

Tested

The repo has 81 Python behavior tests (import edge cases, damaged transcripts, odd paths, parallel saves) plus 6 tests for the mod.
GitHub Actions runs them on Linux and macOS with Python 3.9 and 3.13 on every push and pull request.

Try it

If you hit a bug, especially on Windows (best effort, untested), open an issue.
I would like to know what breaks.

Top comments (0)