# stillyou

English | [中文](./README.zh-CN.md)

**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.

```console
$ 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)

```bash
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](./docs/guide.md#cross-host-mcp). Building from source needs [bun](https://bun.sh): `bun install && bun build --compile packages/cli/src/main.ts --outfile ~/.stillyou/bin/stillyou`.

## The five commands you'll actually use

| | |
|---|---|
| `stillyou search <keywords>` | find past conclusions, files, methodology (superseded versions hidden) |
| `stillyou get <id>` | full entry + provenance: which session, from what inputs, why |
| `stillyou verify <id>` | endorse what you've checked — it ranks up and never decays to zero |
| `stillyou status` | ledger health: entry breakdown, stale warnings, failed-distill alerts |
| `stillyou schedule --at 04:00` | prefer one nightly batch over per-session distillation |

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

| Versus | The one-line boundary |
|---|---|
| Just re-running | Re-runs can't buy back **consistency** or **verified status** — no model speed fixes two reports that disagree |
| claude-mem & memory frameworks | They record the **process** (what happened); stillyou records the **results** (what remains, what superseded what, what's verified) |
| CLAUDE.md / platform memory | Platforms remember *facts about you*, locked in their fence; stillyou records *your outputs*, in plain files you can walk away with |
| git | stillyou's territory is git's blind spot: decision rationale, rejected alternatives, outputs that never enter a repo |

**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](./docs/工作流候选指南.md) and [use tracking](./docs/工作流使用指南.md) (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](./docs/评测指南.md) (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](./docs/工作流发布指南.md) (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](./docs/guide.md)** — data format, cost & privacy (read this before cloud-syncing the ledger), scopes, MCP, architecture, roadmap
- [Design rationale](./docs/2026-07-15-agent-km-产品调研与设计.md) (Chinese) · [A usage story](./docs/我如何用akm.md) (Chinese)
- [The long game: externalized judgment](./docs/guide.md#the-longer-you-use-it-the-more-it-sounds-like-you) — what the ledger becomes after months of use

MIT © An-idd