Build a Memory Assistant with Cria

A minimal demo of a chat assistant that remembers users across sessions with Cria.

• Single-page App Router UI
• Streaming chat API route
• SQLite-backed summaries + vector recall
• Explicit Save session step so memory boundaries are visible

---

Quick start

npm install
cp .env.example .env

Add your key in .env:

OPENAI_API_KEY=your_key_here

Run:

npm run dev

Open http://localhost:3000.

---

Demo flow

1. Send a personal message, e.g.
   • Hi, I'm Alex. I work as a data scientist and I love TensorFlow.
2. Ask one more related question.
3. Click Save session.
4. Ask:
   • What do you remember about me?

You should see recall from the saved session.

---

How memory works

On each chat request, the prompt is composed from:

1. App system prompt
2. Long-term user summary
3. Vector recall from prior saved sessions
4. Session summary for older turns (when needed)
5. Recent messages verbatim

When you click Save session, we:

• summarize the current session,
• update the rolling user summary,
• index the session summary in the vector store,
• clear in-memory session state.

SQLite file: data/cria.sqlite

---

Read code in this order

1. src/lib/memory.ts — memory composition + persistence
2. app/api/chat/route.ts — prompt rendering + streaming response
3. app/api/reset/route.ts — explicit session finalization
4. app/page.tsx — minimal demo UI

---

Environment variables

Variable	Default	Purpose
OPENAI_API_KEY	required	OpenAI auth
LLM_MODEL	gpt-5-nano	Chat model
TOKEN_BUDGET	40000	Prompt render budget
MAX_RESPONSE_TOKENS	700	Max output tokens
MEMORY_VECTOR_LIMIT	4	Max recalled vector hits
MEMORY_DB_PATH	./data/cria.sqlite	SQLite path

---

Commands

npm run dev         # start dev server
npm run build       # production build
npm run test        # run tests
npm run check       # lint/format checks (Ultracite)
npm run fix         # auto-fix lint/format issues