Adaptive Motivation Protocol

1// Before AMP: The Messy Way
2 
3// Build RAG for user data
4const embeddings = await openai.embeddings.create({
5  input: userHistory,
6  model: "text-embedding-3-small"
7});
8 
9await pinecone.upsert({
10  userId,
11  vectors: embeddings.data
12});
13 
14// Retrieve on every request
15const userContext = await pinecone.query({
16  userId,
17  topK: 10
18});
19 
20// Bloat your system prompt
21const systemPrompt = `
22  User preferences: ${userContext.matches.map(m => m.metadata).join('\n')}
23  User history: ${userHistory}
24  Communication style: ${inferredStyle}
25  ... [2000 tokens of prompt engineering]
26`;
27 
28// Hope it works
29const response = await ai.generate({
30  system: systemPrompt,
31  messages: userQuery
32});
33 
34// Manually track outcomes (maybe?)
35if (userCompleted) {
36  await db.update({userId, completionRate: ...});
37}
38 
39// Total: 300+ lines of code
40// Cost: 2000+ tokens per request
41// Maintenance: Constant prompt tuning

1// After AMP: The Clean Way
2 
3import { AMP } from '@amp-protocol/client';
4const amp = new AMP(process.env.AMP_KEY);
5 
6// Get personalization context
7const context = await amp.getContext({
8  userId: user.id,
9  task: userQuery
10});
11 
12// Adapt your response
13const response = adapt(userQuery, context);
14 
15// Report outcome (AMP learns automatically)
16await amp.reportOutcome({
17  requestId: context.requestId,
18  started: true,
19  completed: true
20});
21 
22// Total: 3 lines of code
23// Cost: Minimal tokens
24// Maintenance: Zero