llm011

Ethan Agent — AI skill for Claude Code

AI community

A lightweight, high-performance personal AI agent.

How to install Ethan Agent

This entry records only its repository, not the path inside it, so there is no exact command to give. Open llm011/ethan-agent and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Ethan Agent does

A lightweight, high-performance personal AI agent. Fast async core, pluggable tools, and open integration with external agent ecosystems (Claude Code, Codex, MCP). Simple to configure, easy to extend.

Alternatives in AI

Deepreasoning — A high-performance LLM inference API and Chat UI that integrates DeepSeek R1's CoT reasoning traces with Anthr 5.4k ★
Vllm Mlx — High-performance OpenAI and Anthropic compatible LLM inference server for Apple Silicon 1.5k ★
Easy LLM CLI — An open-source AI agent that is compatible with multiple LLM models 1.3k ★

README

Ethan Agent

[中文文档](./README_CN.md)

A lightweight, extensible personal AI agent built in Python. Designed to run persistently on your own hardware with memory that grows over time, scheduled tasks, and a pluggable tool/skill system.

Ethan combines ideas from [OpenClaw](https://github.com/openclaw/openclaw) (structured agent loop, layered memory), [Hermes Agent](https://github.com/NousResearch/hermes-agent) (self-improving skills, memory consolidation), and [nanobot](https://github.com/HKUDS/nanobot) (minimal core, readable codebase).

Features

**Memory system**

Structured long-term memory (memory.db, the single source of truth for user facts): short sessions (<3 turns) extract immediately, then incrementally every 3 turns — each candidate carries an exact quote from a user message — and is deterministically admitted — explicit → active immediately, observed → promoted only after ≥2 independent sessions. 64 typed dimensions across 7 categories (personal info / preference / activity / decision / relationship / methodology / companion), with TTL expiry, supersede chains, and redaction-on-forget
Semantic recall & dedup: hybrid FTS5 + BGE vector retrieval (RRF-fused) feeds a single prompt block; at admission, embedding near-neighbors are paired and merged/superseded by deterministic rules ("住在深圳" ≈ "家在深圳南山" never stored twice)
Memory decay & reinforcement: fully deterministic tier rules (zero LLM) — identity/preference/companion memories are exempt from decay; project scopes idle for 21 days are archived (dormant, restorable at any time); tentative decisions ("先试试") sink on a 3-day ranking half-life; evidence repeated across independent sessions promotes confidence up a ladder (2→0.8, 3→0.9, 5→0.95). Value is voted by future behavior: recurrence reinforces, long silence archives
Dimension registry: the extraction prompt's dimension guide and the validation whitelist are both generated from one declarative registry