
Epiphany
io.github.GATSV-tech由 GitHub 账号 GATSV-tech 发布v0.5.0更新于 Oct 25, 2026
One private, local memory your AI tools share. Cross-model, on-device, never lose a session.
概览
Epiphany 是一个本地私有的记忆服务器,让 AI 助手在本机不同 AI 工具之间共享并调用上下文记忆。
- 可以做什么
- 从本地共享记忆图谱中召回已存的想法及其推理依据,让助手接续此前会话或其他 AI 工具的上下文。保存新记忆,并在想法之间建立明确的推理连接,例如 because、builds_on。通过后台监视器读取已连接工具的会话记录,自动归档已结束甚至崩溃的会话。通过仪表盘和交互式图谱查看、搜索和整理记忆。将服务器接入受支持的 AI 工具,并安装常驻规则,让智能体主动使用记忆。
- 适用场景
- 当你同时使用多个 AI 工具、反复解释相同背景,或希望聊天结束后仍保留推理过程时适用。也适合希望记忆完全留在本机、不依赖云端账户的用户。
- 运行要求
- 在桌面机器上以本地进程运行,需要 Node.js 18 或更高版本,可通过 npm 包安装或从源码构建。记忆保存在用户主目录下的 SQLite 文件中。无需账户、API 密钥或云服务;可选的 weave 命令会使用你自己的 LLM 密钥。
安装
在 SourceWeft 中
- 打开 控制台中的 Epiphany,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
Epiphany
One private, local memory your AI tools share.
Every AI you use forgets you the moment the chat ends. You re-explain the same context over and over, and when a session dies, the reasoning behind it is gone for good. Epiphany fixes that. It's a single memory that lives on your machine, that Claude, Grok, Gemini, and more all read from and write to, so what you tell one, the others already know.
- Local-first. Your brain is a file on your disk. Nothing leaves the machine. No account, no cloud, no company owns your memory.
- Cross-model. One shared memory across Claude (Desktop + Code), Grok, Gemini, Cursor, Antigravity, Warp, Hermes, and Perplexity. Tell Grok something, ask Claude about it.
- Real memory, not bullet points. Thoughts carry their reasoning and connect into a graph, so recall returns why something is true and what it followed from, not a lone fact. Ask a follow-up and it holds the thread.
- Never lose a session. A background watcher reads the transcripts your tools already write to disk and files every session into the brain in near-real-time, even ones that crashed before anything was saved.
How it works
Epiphany is a local MCP server plus a tiny CLI.
- It runs a small server on your machine that exposes two tools to any AI client:
recallandremember. epiphany connect <tool>wires that server into each AI tool, in the tool's own config format, and installs a standing rule so the agent reaches for your memory on its own.- Memories are embedded on-device (a local model, nothing sent anywhere) and stored in SQLite. Agents wire explicit reasoning connections between thoughts (
because,builds_on,part_of, ...), and recall walks that graph to reconstruct the whole thought-string. - A background watcher sees your tools' session logs the moment they change and sweeps them into the brain, so nothing is ever lost.
The intelligence stays in the models you already use. The brain itself is a fast, private librarian: it holds, connects, and recalls. It never phones home.
Quickstart
Requires Node 18+.
Or build from source:
Wire it into the tools you use, then fully restart each one:
Turn on the background "never lose a session" sweep:
That's it. Now talk to any connected AI. It will recall what you've told the others and save what's worth keeping, on its own.
Run epiphany with no arguments any time to open the dashboard and look inside your brain.
Supported tools
Commands
Your brain lives at ~/.epiphany/brain.db. It's yours. Back it up, inspect it, delete it, it's just SQLite.
Roadmap
Epiphany's core, local cross-model memory with reasoning graphs and automatic capture, works today. Next:
- More session sources for the ingestor (Claude Code, Cursor transcripts).
- Reconciliation. When a fact changes, the brain flags the conflict and you confirm the update, so memory self-corrects instead of piling up contradictions.
- Cloud bridge. An optional relay so cloud-only surfaces (ChatGPT on the web, wearables) can reach your local brain.
- Desktop app. A download-and-run shell for people who don't live in a terminal.
Contributions and ideas welcome.
License
MIT.
来源:README.md,提交 eecacea
工具
0版本历史
1- v0.5.0最新Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分