Shopping assistants

import { Recall } from "@recall-memory/sdk";

const recall = new Recall({ workspaceId: "shop-eu" });

// The shopper is a peer keyed by your customer id; the assistant is a peer Recall does not
// learn about.
const shopper = recall.peer("cust-1042", { metadata: { country: "NL" } });
const assistant = recall.peer("shop-assistant", { configuration: { observeMe: false } });
// Once, for the whole shop: what Recall should learn here, and what it must not.
await recall.setConfiguration({
  reasoning: {
    customInstructions:
      "Record sizes and fit per brand, budgets, colours, styles, brands, diets the shopper " +
      "states, and returns with their reasons. Never record or guess health, pregnancy, " +
      "religion, ethnicity, sexuality or money trouble from purchases or messages.",
  },
});

// What the shopper entered in their account is fact, not inference: store it as facts.
await shopper.conclusions.create(["Shoe size EU 42.", "Eats vegan."]);

// A return from an earlier order, in the shopper's words from the returns form, dated when
// it was filed. Recall learns from it like any message, and can close facts it contradicts.
const order = recall.session("order-77120");
await order.addMessages([
  shopper.message("Returning the Trailblaze GTX in 42: too small. I need a 43 in this brand.", {
    createdAt: "2026-10-08T17:20:00Z",
    metadata: { order: "77120", sku: "TB-GTX-42", reason: "too_small" },
  }),
]);

// In the shop's chat: one request stores the message and returns what Recall knows about
// the shopper for the reply. Preferences stated here are learned in the background.
const visit = recall.session("chat-5e21c");
const { context } = await visit.turn(
  [shopper.message("Trail shoes for muddy autumn runs, please. Nothing neon, max €150.")],
  { peerTarget: shopper, tokens: 2000 },
);
const prompt = context.toOpenAI(assistant);
// Your model writes the reply from `prompt` and your catalogue; store it too.
const reply = "The Trailblaze GTX in black grips well in mud, €139. In 43 this time?";
await visit.addMessages([assistant.message(reply)]);

// Turn memory into filters for your catalogue: ask for JSON once per visit and keep it in your
// session. (waitForIdle is for this demo; real apps rarely block on learning.)
await recall.waitForIdle();
type Filters = {
  shoe_size_eu: number | null;
  max_price_eur: number | null;
  avoid_colours: string[];
  diet: string[];
};
const prefs = await shopper.chat("Which filters should a product search use for this shopper?", {
  reasoningLevel: "low",
  responseFormat: {
    jsonSchema: {
      type: "object",
      properties: {
        shoe_size_eu: { type: ["number", "null"] },
        max_price_eur: { type: ["number", "null"] },
        avoid_colours: { type: "array", items: { type: "string" } },
        diet: { type: "array", items: { type: "string" } },
      },
      required: ["shoe_size_eu", "max_price_eur", "avoid_colours", "diet"],
    },
    parse: (data: unknown) => data as Filters, // or a zod schema's parse
  },
});
const catalogue = [
  { sku: "TB-GTX-43", size: 43, price: 139, colour: "black" },
  { sku: "VX-9-43", size: 43, price: 165, colour: "neon green" },
];
const picks = catalogue.filter(
  (p) =>
    (!prefs?.max_price_eur || p.price <= prefs.max_price_eur) &&
    !prefs?.avoid_colours.some((c) => p.colour.includes(c)),
);

// The shopper turns personalisation off: stop learning from their messages and drop every
// fact memory holds about them. Their chats stay stored.
// (To erase the shopper completely instead: await shopper.delete())
await shopper.setConfiguration({ observeMe: false });
const learned = await (await shopper.conclusions.list()).toArray();
for (const fact of learned) await shopper.conclusions.delete(fact.id);
await shopper.setCard([]);