Kage7047

Kioku Assistant — AI skill for Claude Code

AI community

Private, self-hosted agentic personal assistant: cloud brain, local hands.

How to install Kioku Assistant

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Kage7047/kioku-assistant and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Kioku Assistant does

Private, self-hosted agentic personal assistant: cloud brain, local hands. Claude Agent SDK over a git-backed markdown vault, with hybrid retrieval, model routing under a budget guard, and a human approval gate.

Alternatives in AI

Hindsight — State-of-the-art long-term memory for AI agents by Vectorize 6.7k ★
Ataru — High-performance AI memory retrieval for local agent history — a Rust search core (SDK / API / JSON CLI) plus 351 ★
Memory Lancedb Pro — Production-grade long-term memory system (v1.1.0-beta.8) for OpenClaw AI agents 222 ★

README

Kioku (記憶)

**A private, self-hosted agentic personal assistant — a cloud brain with local hands.** *Kioku* (記憶, "memory") is a second brain that never sleeps: it owns a markdown knowledge vault, reasons over it with the Claude Agent SDK, works autonomously under a strict budget and approval model, and reaches you through Telegram, a web command center, or an API.

[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE) ![Python](https://img.shields.io/badge/python-3.12+-blue.svg) ![Status](https://img.shields.io/badge/status-reference%20implementation-purple.svg)

![Kioku command center](assets/command-center-desktop.png)

This is a portfolio/reference implementation of a single-user assistant. The design, architecture, and full build log live in [`docs/`](docs/INDEX.md). The owner's private knowledge vault is **not** part of this repository — the code runs against your own.

Why it's interesting

Most "AI assistant" projects are a thin wrapper around a chat endpoint. Kioku is the opposite: a small but complete **agentic system** engineered around three hard constraints — *cost*, *trust*, and *durable memory*.

Files are truth, the database is an index. All knowledge is plain markdown in a git-backed vault. Postgres + pgvector is a rebuildable cache — drop it and lose nothing.
It runs on a subscription, not a metered API. The whole cost architecture (Haiku triage → Sonnet work → Opus orchestration, prompt-cache-friendly prefixes, deterministic Python for anything that doesn't need a model, and a hard budget guard on every run) is designed to keep a genuinely useful always-on assistant inside a personal plan's monthly agent credit.
Autonomy with a human gate. It works end-to-end up to any irreversible or outbound action — sending a message, spending money, deleting outside the vault — which it stages for one-tap approval instead of executing.
It learns you over time. A derived model of you