JavaScript

npm install woolly-ai

Woolly needs to be running: download the Mac app or run woolly serve.

Save and recallCopy Anchor Link

import { Woolly } from 'woolly-ai';

const w = new Woolly({ app: 'My app' });

await w.remember('Moved to Denver in October', { key: 'user.location' });
for (const m of await w.recall('where do I live?')) console.log(m.content);

Give a key for facts that can change. A newer memory with the same key replaces the older one. app is the name your memories are saved under and how your app shows up in woolly log.

Add memory to a promptCopy Anchor Link

const system = `You are a helpful assistant.\n\n${await w.context(userMessage)}`;

context returns the relevant memories as a short note that tells the model they're information, not instructions, or an empty string when nothing is relevant.

Vercel AI SDKCopy Anchor Link

Wrap any model and every call gets the user's relevant memories as a system message:

import { generateText, wrapLanguageModel } from 'ai';
import { woollyMemory } from 'woolly-ai/ai-sdk';

const model = wrapLanguageModel({ model: yourModel, middleware: woollyMemory() });
const { text } = await generateText({ model, prompt: 'Any snack ideas?' });

Options: limit (memories per call, default 6), project, and any new Woolly() option. If Woolly isn't running, calls go through unchanged.

Learn from a conversationCopy Anchor Link

With Ollama running, a local model picks out the facts worth keeping. See Smart memory.

const steps = await w.learn([
  { role: 'user', content: 'I just moved to Denver. Keep answers short.' },
]);

Everything elseCopy Anchor Link

list, get, update, forget, restore, trash, conflicts, resolve and profile. Pass project to keep notes about one codebase separate, see Projects.

Errors are WoollyError, with status set to the HTTP status. When nothing answers at Woolly's address you get WoollyNotRunning.

new Woolly() uses WOOLLY_HOST and WOOLLY_HOME like the rest of Woolly. Requires Node 18 or newer.