Claude Memory

Claude context window vs persistent memory

Claude models support large context windows (200K+ tokens on some models). That helps with long documents in a single session, but it does not replace memory.

When a user returns tomorrow, last week's conversation is gone unless you send it again. Large context replay is costly and still misses facts from older sessions.

Persistent memory stores discrete facts and recalls only what is relevant to the current question.

Claude Projects vs API memory

Claude.ai Projects let consumer users upload files and maintain project-level context. This feature is not available as a per-user memory API for your customers.

If you build a product on the Anthropic API, you own the memory architecture. Databaset provides user-scoped store and recall that works with every Claude model.

Adding memory to Claude API calls

• Recall memories before each messages.create call
• Add recalled facts to the system prompt or a dedicated user context block
• Use Claude's strength at following structured context instructions
• Store new user facts after meaningful exchanges

import Anthropic from "@anthropic-ai/sdk";
import { Databaset } from "@databaset/sdk";

const anthropic = new Anthropic();
const memory = new Databaset({ apiKey: process.env.DATABASET_API_KEY });

async function askClaude(userId: string, question: string) {
  const { memories } = await memory.recall({ userId, query: question, limit: 5 });

  const response = await anthropic.messages.create({
    model: "claude-sonnet-4-20250514",
    max_tokens: 1024,
    system: `You are a helpful assistant. User memories:\n${memories.map(m => "- " + m.text).join("\n")}`,
    messages: [{ role: "user", content: question }],
  });

  return response.content[0].type === "text" ? response.content[0].text : "";
}

Prompting tips for Claude + memory

• Label the memory block clearly: User memories (verified facts)
• Tell Claude to prefer memories over assumptions when they conflict
• Set minScore on recall to filter low-confidence matches
• Use appId to separate staging, production, and EU tenant namespaces

Same memory for Claude and GPT

Databaset is model-agnostic. Store once, recall for Claude, GPT, Gemini, or any LLM. Users keep one memory profile across your entire AI stack.

This is essential for products that let users switch models or run model routing by task type.