Memory Chat — shared trip demo (TripMate)

A small Next.js demo of the @xtraceai/memory SDK paired with the Vercel AI SDK and GPT-4o-mini.

TripMate is a collaborative trip-planning assistant. Two travelers — Alice and Bob — are planning one Tokyo trip. The assistant answers each of them using three layers of memory at once, blended into a single prompt by one recall() call:

Layer	Scope axis	Example
Personal	user_id	Alice is vegetarian and likes boutique hotels; Bob is a budget-minded ramen fan
Shared trip	group_ids	"May 10–17", "staying near Shibuya", "one group sushi dinner" — visible to everyone on the trip
Travel guide	app_id	"buy the JR Pass before arrival", "tipping isn't customary", "Afuri = yuzu ramen" — seeded by the product, read by all

Switch traveler in the header and watch the personal slice change while the trip and guide stay shared. The Context tab shows the exact assembled prompt, sectioned by source, with author attribution (you: / alice: / bob:).

The point

recall() is the combined read that plain search() can't express, because search AND-narrows every axis you pass. Each turn:

1. Recall unions three independent scope pools — personal OR trip OR guide — dedupes, score-ranks, and renders one source-sectioned context block.
2. Call the LLM with that block injected into the system prompt.
3. Ingest the turn back under the traveler's user_id and offer it to the trip group — the ingest classifier tags only the trip-relevant facts. The guide (app_id) is never written from chat.

// app/api/chat/route.ts — one recall unions three scopes
const { prompt } = await client.memories.recall({
  query,
  pools: [
    { user_id: persona },           // this traveler's own prefs
    { group_ids: [tripGroupId] },   // the trip's shared facts (cross-user)
    { app_id: 'tripmate-guide' },   // the provider travel knowledge base
  ],
});
// axes AND within a pool; the three pools OR. The guide rows render under their
// own section, not framed as the user's memory.

Run locally

cp .env.example .env.local
# Fill in XTRACE_API_KEY, XTRACE_ORG_ID, OPENAI_API_KEY (XTRACE_BASE_URL optional)

npm install
npm run dev
# → http://localhost:3000

Then click Seed demo once to register the trip group and load Alice/Bob/trip/guide memories, and start asking questions like "where should we grab dinner near the hotel the first night?"

How the pieces fit

File	Role
lib/personas.ts	Persona constants (Alice/Bob) — SDK-free so the client bundle stays light.
lib/memory.ts	Lazy MemoryClient; trip-group + product-KB constants; getTripGroupId() find-or-create helper.
app/api/seed/route.ts	One-shot seed: registers the trip group and ingests personal / trip / guide memories. Idempotent.
app/api/chat/route.ts	POST chat. recall() (3 pools) → inject → streamText → onFinish ingest (user + group).
app/api/recall/route.ts	Returns the assembled recall() prompt + per-scope counts for the Context panel.
app/api/memories/route.ts	GET list/search scoped to the selected traveler; DELETE wipes both travelers (guide stays).
app/page.tsx	Chat UI, persona switcher, and the tabbed sidebar (Context / Memories).

How memory is wired

The recall() call above is the whole story: one request, three scopes, one prompt. Group resolution (grp_… → name) and source-sectioning happen inside the SDK. Tagging is best-effort at ingest — the classifier decides which extracted facts belong to the trip group based on the group's prompt.

The @xtraceai/memory/ai-sdk subpath also ships createXtraceMemory (a "memory just works" model wrapper) and memoryTools (tool-calling). This demo calls recall() directly so the three-scope assembly is visible — but memoryTools accepts group_ids if you'd rather let the LLM drive.

Deploy

Pushes to main auto-deploy via Vercel. Set the env vars in Project Settings → Environment Variables:

• XTRACE_API_KEY
• XTRACE_ORG_ID
• OPENAI_API_KEY
• (optional) XTRACE_BASE_URL, DEMO_CONV_ID

What this demo deliberately doesn't have

• Auth / login — the two travelers are hardcoded personas, not real accounts. Knowing a group id is the access boundary; a real app maps its own users to user_ids and decides which group_ids to send.
• Per-traveler chat history across reloads (memory persists; the chat thread doesn't — that's the point).
• Tests, and streaming-UI polish beyond the Vercel AI SDK defaults.