
Woolly 推出跨 AI 工具的共享本地记忆
来源标题:Show HN: Woolly, a local mem0-compatible memory API for AI agents
Woolly 在 Hacker News 的 Show HN 板块发布,主打让所有 AI 工具共用同一份本地记忆。该项目面向需要跨工具保留上下文的用户,具体实现与支持范围未在标题中说明。
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
memory and learning for your agents
Woolly is the open memory engine for agents. It keeps facts current and learns from every session, through an API, Python and JavaScript SDKs, plugins and MCP. It runs on your own machine by default, in Markdown you can read.
Get startedDownload for Mac
Works with
Why Woolly
Every AI coding tool starts from zero. You explain your stack to Cursor, then again to Claude Code, then to Codex. Each keeps its own memory, or none, and none of them know what you decided yesterday or what your teammate’s agent learned this morning.
Woolly is the memory they share. It’s one folder on your laptop, plain Markdown, with no cloud and no account. Any tool that speaks MCP can save to it and search it, and woolly proxy gives the same memory to Ollama and any app that uses the OpenAI or Anthropic API.
When a session starts, your agent gets your profile, the project’s notes and where you left off. When it ends, Woolly saves what you decided and why.
The agents you build get the same memory. Send a conversation to the memory API and a local model decides what’s new, what changed and what’s no longer true, per user. While you’re away, Woolly dreams. It tidies what it knows and learns lessons from your sessions.
Sessions that start warm.
Claude Code opens every session with your profile, the project’s notes and a handoff from last time. Nobody has to say “here’s my stack” again.
Decisions stick.
When a session ends, Woolly saves what you chose, why, and what you ruled out. No agent pitches the rejected option again.
Agents that remember.
Give the agents you build a memory with an API shaped like mem0’s. For each fact a local model decides whether it’s new, an update, or no longer true, and every user’s memories stay apart.
It dreams.
While you’re away, Woolly merges duplicates, settles facts that changed, and turns your past sessions into lessons for the next one.
Facts keep their history.
A newer fact replaces the old one, which stays on disk with the time it stopped being true. When an AI guesses something that contradicts you, your version wins and the guess waits for your review.
Memory you can read
Every memory is a Markdown file. Open it, grep it, fix it by hand, and see which tool saved it. The Woolly app shows the same thing with search.
One project memory for the team.
woolly init keeps the repo’s stack, conventions and gotchas in .woolly/woolly.md. Commit it, and every teammate’s agent reads and updates the same notes.
Build agents that remember
The same calls as mem0, answered by your own machine. Install the package, add a conversation for a user, and search before your agent replies.
import { Woolly } from 'woolly-ai';
const w = new Woolly();
await w.add('I just moved to Austin', { userId: 'alice' }); // [{ memory: 'Lives in Austin', event: 'UPDATE', previous_memory: 'Lives in Denver' }]
const memories = await w.search('where does she live?', { userId: 'alice' });
Memory APIJavaScriptPython
Our promise is memory you can check, not a number you have to trust.
382 questions over the first three conversations, searched on a laptop with nomic-embed-text. No hosted model and no API key. Run the same test yourself.
Woolly runs where your code does, with no cloud at all.
It’s one binary with no outside dependencies. Install it on your laptop, sync it to your other machines, or commit the project’s memory so your team shares it.
Mac, Windows and Linux. One small binary, or the app.
Sync your machines over your own network, encrypted.
Project memory in git, shared with your whole team.
Woolly refuses passwords and API keys, and an AI can’t plant instructions in your memory. Lines that come from a repo are checked for secrets and injected instructions before any agent sees them.
Read the code
Shipping weekly
Built in public, one release at a time. Here’s what landed lately.
v0.14.1Nov 17Search by meaning is more accurate. Woolly now gives nomic-embed-text the search and document prefixes it was trained with (and the query prefix mxbai-embed-large expects).
v0.14.0Nov 17A memory layer API for the agents you build, with mem0's routes and field names: POST /v1/memories turns a conversation into memories, and POST /v1/search finds them.
v0.13.1Nov 13A Windows app. Download Woolly-windows.exe, run it, and Woolly starts in the background with an icon in the system tray.
v0.13.0Nov 13Project memory in git: woolly init. A repo's memory lives in .woolly/woolly.md, one fact per line under Overview, Stack, Conventions, Decisions, Gotchas and Commands, so every teammate's agent reads and updates the same notes.
Read the changelog
Questions
Does Woolly send my data anywhere?No. There are no servers, no accounts and no telemetry. Your memories live in ~/.woolly on your laptop, and that is the whole story.
Which apps work with Woolly?Anything that speaks MCP. Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Codex, Gemini CLI, LM Studio and opencode are one woolly connect away. Ollama and other local models get memory through woolly proxy.
Do I need Ollama?No. Woolly works on its own. With Ollama running, it also searches by meaning, learns facts from your chats and saves decisions when a session ends.
How is this different from ChatGPT or Claude memory?Those memories are locked inside one app, on that company's servers. Claude never hears what you told Cursor. Woolly is one memory every tool shares, and it lives on your laptop in files you can read.
What happens when a fact changes?A new memory with the same key replaces the old one. If an AI saves something that disagrees with what you said, your version stays and the AI's is saved as a conflict for you to review.
Can my team share memory?Yes. Run woolly init in a repo and commit .woolly/. Every teammate's agent reads and updates the same project memory, and facts about you never go in the repo.
Can I use Woolly as the memory for agents I build?Yes. The memory API uses mem0’s routes and field names, with JavaScript and Python packages. Add a conversation for a user, then search before your agent replies. It runs on your own machine and only answers that machine, so it fits agents you and your team run, not a hosted service for thousands of users.
How is Woolly different from mem0 or supermemory?Woolly is one open-source program that runs on your own machine, with your memories in Markdown files you can read. Its API is shaped like mem0’s, so switching is mostly a new address. It keeps every old version of a fact, never lets a model overrule what you said, and dreams to turn your sessions into lessons.
Is Woolly free?Yes. It is MIT licensed and built in public on GitHub. Issues and pull requests are welcome.
Still have a question? Ask on GitHub.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。