stillyou

English | 中文

Memory tools remember what your agent did. stillyou remembers what you decided — and what actually shipped.

Your agent forgets. What remains is still you. — distill yourself, one session at a time.

Every session ends in amnesia: the deliverables, the decisions, the rejected alternatives, the methodology you settled on — gone. stillyou captures them automatically and hands them to your next session.

$ claude
> Continue the AI video tool comparison from last time

⏺ Your ledger has relevant history:
  [5802983d] video-tool-pricing-comparison@v1 (file/final, verified)
  Last conclusion: Kling wins on cost-per-second (¥66/mo ≈ 66s of video).
  Building the incremental comparison on that methodology — not re-deriving it...

A fresh session, zero re-explaining. That answer cites a ledger entry distilled from a session three weeks ago — same methodology, human-verified conclusion.

Install (30 seconds)

curl -fsSL https://raw.githubusercontent.com/An-idd/stillyou/main/install.sh | sh
~/.stillyou/bin/stillyou init        # pick a ledger location + register Claude Code hooks — done

Works with Claude Code today; Codex / ZCode / Cowork read the same ledger via MCP. Building from source needs bun: bun install && bun build --compile packages/cli/src/main.ts --outfile ~/.stillyou/bin/stillyou.

Related MCP server: Agent Directory MCP Server

The five commands you'll actually use

Everything else is automatic: writes are journaled by hooks (zero LLM), sessions distill in a detached background process, and your first message each session pulls in relevant history under a hard token budget.

How it differs from what you already have

Built for: solo creators, serial writers, analysts, vibe coders — anyone whose work is a series, not one-offs, and whose outputs live outside git. Not for: one-off tasks, chat-only use, heavyweight engineering teams (git + CI is their home turf).

Three promises: your data stays in plain markdown + jsonl you can grep (delete stillyou, keep everything); hooks never block or break your sessions; uncertain provenance stays blank — no guessing, ever.

Learn more

• Experimental workflow learning and use tracking (Chinese) — propose methods from task and correction evidence, select by scope and declared prerequisites through CLI/MCP, and trace adoption reports, tool evidence and task reviews. No automatic injection or proven performance gains yet.

• Workflow evaluation (Chinese) — bind an exact workflow revision to a frozen A–D replay, inspect paired task results, and diagnose automatic selection against an explicit reference using the versions actually offered. Selection quality and synthetic fixtures remain separate from real task effects.

• Workflow releases (Chinese) — explicitly pin a version and its selection conditions while developing a new candidate; switch, roll back or unpublish with an auditable history. Manual publication does not mark a workflow verified.

• Full guide — data format, cost & privacy (read this before cloud-syncing the ledger), scopes, MCP, architecture, roadmap

• Design rationale (Chinese) · A usage story (Chinese)

• The long game: externalized judgment — what the ledger becomes after months of use

MIT © An-idd