Agent Memory Layer — Development skill for Claude Code
Local-first shared memory + skill governance for multiple coding agents (Claude Code / Kimi Code / DeepSeek Harness).
How to install Agent Memory Layer
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open kh859278/agent-memory-layer and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Agent Memory Layer does
Local-first shared memory + skill governance for multiple coding agents (Claude Code / Kimi Code / DeepSeek Harness). MCP-ready, no cloud.
Alternatives in Development
- IPolloWork — Enterprise-grade, local-first Agent Workbench for people and agent teams 4.9k ★
- Alethe Agents — A local-first desktop workspace for running, organizing, and resuming multiple coding agents and shells with r 511 ★
- Flowix — Notes for you, Memory for your agents 374 ★
README
agent-memory-layer
[](https://github.com/kh859278/agent-memory-layer/actions/workflows/ci.yml) [](https://pypi.org/project/aml-memory/) [](pyproject.toml) [](LICENSE)
**Multiple coding agents. One shared memory. One skill state.** Claude Code / Kimi Code / DeepSeek Harness side by side — they can share what each of them learned, **fully local**, no cloud, no daemon.
**多个 coding agent,共用一份记忆与技能状态。** 把 agent 历史从"各自锁在目录里的流水账" 变成**可复用、可治理、可验证**的资产。
| 环节 | 做什么 |
|---|---|
| 采集 | 从各 agent 的会话文件挖历史,保留原始时间戳(不是入库时间——否则时序全假) |
| 沉淀 | 会话 → 跨项目知识(带复核期),并同步一份人能读的 markdown(不被锁进数据库) |
| 检索 | 分阶段 P0–P6(每阶段有固定条数/字数预算)、级联回退、未命中必须解释"为什么空" |
| 治理 | 技能上游监控 + 能力声明(skill.yaml)+ 9 态生命周期:未批准的版本只暂存,不自动生效 |
| 验证 | 检索层基准(问法对照 / P@3 / 注入量分位数)+ 任务级基准(真起 agent 判成败、算成本) |
**5 分钟跑通** → [`docs/QUICKSTART.md`](docs/QUICKSTART.md) · **常见问题** → [`docs/FAQ.md`](docs/FAQ.md) · **为什么这么设计** → [`docs/PROTOCOL.md`](docs/PROTOCOL.md)
**状态:alpha。** 接口还会变。已发布到 PyPI:**`pipx install aml-memory`**(命令是 `aml`)。 注意 PyPI 上的 `agent-memory-layer` 是**另一个不相关的项目**,只是名字和本仓库重了。 测试与 CI 见上方徽章。 测试与 CI 见上方徽章。
为什么需要它
你同时用 DSH / Claude Code / Codex / Kimi 这类 coding agent。每个 agent 的历史都锁在自己的目录里: 换个 agent 就"不记得"上次怎么解决的;而且这些历史只是流水账—— 第二次遇到同一个坑,还得重新踩一遍。
`aml` 把"历史"变成"资产":
| 做的事 | 具体行为 |
|---|---|
| 采集 | 从各 agent 的会话文件里挖对话(一条"任务"一条"回复"),保留原始时间戳而不是写入时间 |
| 蒸馏 | 会话结束/归档时用 LLM 提炼成跨项目可复用的知识(kind:knowledge + domain:<领域>),与"只对本项目成立"的流水账分开存 |
| 检索 | 分阶段检索协议 P0–P6:每阶段有固定条数与字数预算、层级优先(跨项目沉淀 > 本项目历史 > 全库)、同查询去重;未命中必须解释为什么空 |
| 人面镜像 | 沉淀同步落成 knowledge/沉淀/<领域>.md + 自动重建索引:没有本程序也能读、能 grep、能搬走 |
| 技能治理 | 你装的 skill 也能被纳管:比上游版本、干净就自动更新、改过就只暂存(见下) |
30 秒看效果
体检一条命令(输出为格式示意,真实运行时
Related Skills
Recallnest
One memory, every terminal. Shared memory layer for any MCP client — Claude Code, Codex, Kimi, Antigravity (ag
Tokrs
A command-line interface (CLI) for analyzing token usage in local logs, supporting Claude, Codex, OpenCode, Ki
Unified Agent Memory
Unified fleet-wide agent memory system for DeepSeek Harness — shared Obsidian Vault for dsh, Codex, Claude Cod
/amux — Amux Session Integration
You are running inside an amux managed Claude Code session. amux is a local multiplexer that manages multiple
Dsh Usage
🌊 Persistent dock & fully-customizable balance/usage panel for DeepSeek Harness — activity heatmap, dual-chan
Agentkib
Local-first agent asset manager for Codex, Claude Code, DeepSeek Harness, OpenClaw, Hermes and Cursor — one kn
Related Agents
Harness Eng
Use for every build task on a deepseek-harness core — one layer per claimed packet, gated by hedgehog verify.
Kimi Worker
Forward a coding, diagnosis, or review task to Kimi K3 (Moonshot) through the local kimi-worker.sh runtime. Us
ChatGPT On Wechat
CowAgent是基于大模型的超级AI助理,能主动思考和任务规划、访问操作系统和外部资源、创造和执行Skills、拥有长期记忆并不断成长,比OpenClaw更轻量和便捷。同时支持微信、飞书、钉钉、企微、QQ、公众号、网页