Never Ask Twice

Support that remembers.

Enterprise Support MemoryAgent on Qwen Cloud

Live demo: neverasktwice.dev — try /chat, or hit /health directly.

Customers don't want a smarter chatbot if they still have to repeat their SLA, setup, open issue, and escalation contact every time they come back.

Never Ask Twice is a production-shaped B2B support memory agent that remembers customer context across sessions, retrieves only the memories that matter, forgets stale facts safely, and proves improvement with a deterministic memory ON/OFF evaluation harness plus a live Qwen-backed API path.

The demo agent is Nat. Nat is powered by NATE — the Never Ask Twice Engine — a scoped memory layer that turns support conversations into durable, auditable customer context.

Built for the Qwen Cloud Global AI Hackathon — Track: MemoryAgent.

Brand assets

Official logo files and usage rules live in docs/assets/brand. The public tagline is Support that remembers. Descriptor: Enterprise Support MemoryAgent.

Related MCP server: Memsolus MCP Server

Status

Judge path

1. Try the live deployment: neverasktwice.dev — /chat for the UI, /health for capability status.

2. Read the memory model: docs/memory-model.md.

3. Run the ablation:

   pnpm eval

4. Inspect the forgetting behavior: docs/forgetting-policy.md.

5. Read the system architecture: docs/architecture.md.

6. List MCP tools:

   pnpm build
   pnpm mcp:list-tools

7. Review deployment proof: deploy/railway.md (live, primary) and deploy/alibaba-fc.md (secondary target, wired and ready).

8. Read the build log: Building customer support memory that survives an audit.

9. Watch the demo: https://youtu.be/P254DPj-Mgw.

The measurable result

Run:

pnpm eval

Expected deterministic fixture output:

memory-on re-ask rate: 0.00
memory-on recall accuracy: 1.00
memory-off re-ask rate: 1.00
memory-off recall accuracy: 0.00
re-ask rate: 0.00 (memory) vs 1.00 (no-memory)

The evaluation path is intentionally deterministic for reproducible scoring. It uses fixed synthetic fixtures and a fake Qwen client. The live API path uses Qwen Cloud when DASHSCOPE_API_KEY is configured.

What makes it a MemoryAgent

Never Ask Twice implements explicit memory tiers:

• Working memory — current-session context usable before session-close distillation.

• Episodic memory — raw support events with Qwen embeddings and provenance.

• Semantic memory — distilled customer facts with confidence, validity windows, and source links.

• Forgetting policy — TTL expiry, supersession, stale-memory exclusion, and audit-safe provenance.

• Budgeted recall — relevant memories only, capped to a strict context budget.

• MCP surface — memory tools exposed for agent interoperability.

This is not transcript logging. It is structured memory with retrieval discipline, provenance, forgetting, and measurable cross-session improvement.

Architecture

Customer chat / MCP
        |
        v
Hono API (same code, two deploy targets:
           Railway - primary, live | Alibaba FC - swap-to-when-cleared)
        |
        v
MemoryService
  |-- working memory: current-session facts
  |-- episodic memory: turn events + Qwen embeddings
  |-- semantic memory: distilled durable facts
  |-- forgetting: TTL + supersession + scoped recall
        |
        +--> Qwen Cloud via DashScope-compatible OpenAI API
        +--> Neon Postgres + pgvector
        +--> MCP stdio tools

Full diagram and component map: docs/architecture.md.

Getting started

1. Clone and install

git clone https://github.com/marcelle-labs/never-ask-twice.git
cd never-ask-twice
pnpm install

2. Configure environment

cp .env.example .env

Edit .env and set your DASHSCOPE_API_KEY from DashScope for live Qwen-backed embeddings, distillation, and adjudication. The example is pre-filled for local Postgres on port 5433.

DATABASE_URL=postgresql://neverasktwice:neverasktwice@localhost:5433/neverasktwice
DASHSCOPE_API_KEY=your-key-here
QWEN_BASE_URL=https://dashscope-intl.aliyuncs.com/compatible-mode/v1
QWEN_CHAT_MODEL=qwen-plus
QWEN_EMBEDDING_MODEL=text-embedding-v3
QWEN_EMBEDDING_DIM=1024
MEMORY_TOKEN_BUDGET=1200

Without DASHSCOPE_API_KEY, the API boots in local-safe mode. Local-safe mode uses zero-vector embeddings and empty distillation responses so the server can run without secrets; it does not perform real Qwen work. Use pnpm eval for deterministic local scoring without a key.

3. Start Postgres

docker compose up -d

The local database binds to localhost:5433 so it does not collide with other Postgres services on 5432.

4. Run migrations

pnpm migrate

5. Run the eval harness

pnpm eval

6. Run the boundary scan

pnpm boundary-scan

7. Start the local API

pnpm dev

The API will be available at http://localhost:3000 with endpoints:

• GET /health — health and capability status.

• POST /turn — append a customer/agent turn.

• POST /sessions/:id/close — close a session and distill episodic memory into semantic memory.

• POST /recall — recall a bounded memory bundle.

8. Run the MCP server

pnpm build
node dist/src/mcp/server.js

The MCP server exposes four tools: recall_memory, write_memory, distill_session, and forget.

Project structure

• apps/api — Hono API, local server, and Function Compute handler. Deploy-target-agnostic; live on Railway today.

• src/agent — deterministic support-agent policy used by the eval harness.

• src/contracts.ts — memory predicate enum, Zod contracts, and shared types.

• src/db — Drizzle schema and SQL migration string.

• src/memory — memory service, stores, retrieval, supersession, and forgetting behavior.

• src/mcp — stdio MCP surface over the shared memory service.

• src/qwen — single Qwen Cloud client module.

• src/testing — deterministic fake Qwen client for the eval harness.

• eval — frozen three-session scenario, ground truth, expected output, and runner.

• scripts — boundary scan, migration, MCP list-tools, and demo script checks.

• docs — judge-facing architecture, memory model, evaluation, and forgetting documentation.

• deploy — Railway deployment proof (live, primary) and Alibaba Function Compute deployment instructions (secondary target, wired and ready).

Key commands

Security and clean-room boundary

Never Ask Twice uses synthetic data only. Do not commit real customer data, secrets, .env files, or private platform identifiers. The repository includes a boundary scan to fail on known forbidden tokens and a local-safe mode so judges can run the server without secrets.

See SECURITY.md.

License

Apache-2.0