v0.1For Claude Code power users

Every Claude Code session, filed.

A tiny GUI that watches your past terminal transcripts, clusters them into topics, and opens a fresh claude tab pre-briefed on the ones you click.

$curl -fsSL dossierkit.com/install.sh | sh

macOS onlyTerminal.app + iTerm2Requires Claude Code CLILocal · no telemetry

How it works

One Python script, one zsh hook, one static page. Nothing leaves your machine.

01 · INDEX

TF-IDF over the JSONL session files in ~/.claude/projects/. Topics cluster by frequency and recency. No vector DB.

~/.claude/projects/ -Users-you/ 012a4f-…jsonl 037bcd-…jsonl …

02 · OPEN

A zsh hook (debounced) opens the wordcloud in your browser the first time you open a Terminal tab. Subsequent tabs refocus the existing window.

# appended to ~/.zshrc source ~/dossierkit/src/hook.sh

03 · RESUME

A new Terminal tab spawns running claude with the top 8 past sessions on that topic attached as context. You pick up where you left off.

$ claude "Resume topic 'naban'. Read these transcripts: …"

That's the whole product.

Every transcript under ~/.claude/projects/, scored by TF-IDF, surfaced as clickable chips.

A new claude tab opens with the top 8 past sessions on that topic attached. You skip the blank prompt.

A zsh hook fires on new Terminal tabs. Debounced — won't spam if you open ten tabs in a minute.

Transcripts stay on your machine. The wordcloud is served by a local Python HTTP server. No account, no telemetry, no cloud index.

Three honest options for "don't make me re-explain my project to Claude."

Tool	What it does	Where it runs	Price
DossierKit	Topic wordcloud over your past CC sessions, click → fresh briefed tab.	Local — Python + zsh hook	Free · source-available
claude-mem	Vector-DB memory layer for Claude Code, MCP-style.	Local — Chroma + uvx	Free · open source
mem0 / Letta	Generic LLM memory APIs (cross-tool, not CC-specific).	Hosted	$0–$249 / mo