Mem0 Vercel AI SDK Provider

Memory-enhanced AI provider for Vercel AI SDK. Automatically retrieves and stores memories during LLM calls.

Step 1: Install

npm install @mem0/vercel-ai-provider ai@^6

ai and the @ai-sdk/* provider packages ship as regular dependencies of @mem0/vercel-ai-provider, so nothing else needs installing. The only peer dependency is zod (optional, ^3.0.0).

Step 2: Set up environment variables

export MEM0_API_KEY="m0-xxx"
export OPENAI_API_KEY="sk-xxx"   # or ANTHROPIC_API_KEY, GOOGLE_API_KEY, etc.

Get a Mem0 API key at: https://app.mem0.ai/dashboard/api-keys?utm_source=oss&utm_medium=skill-mem0-vercel-ai-sdk

Pattern 1: Wrapped Model

The wrapped model approach is the simplest. createMem0 returns a provider that wraps any supported LLM with automatic memory retrieval and storage.

import { generateText } from "ai";
import { createMem0 } from "@mem0/vercel-ai-provider";

const mem0 = createMem0();
const { text } = await generateText({
  model: mem0("gpt-5-mini", { user_id: "alice" }),
  prompt: "Recommend a restaurant",
});

What happens under the hood:

1. The prompt is stored to Mem0 (POST /v3/memories/add/), awaited before the LLM call (a failed write is logged and ignored)
2. The prompt is sent to Mem0 search (POST /v3/memories/search/) to retrieve relevant memories
3. If any memories are found, they are injected as a system message at the start of the prompt
4. The underlying LLM (e.g., OpenAI gpt-5-mini) generates a response using the enriched prompt

For generateText, the retrieved memories are also attached to the result as a source:

const { text, sources } = await generateText({
  model: mem0("gpt-5-mini", { user_id: "alice" }),
  prompt: "Recommend a restaurant",
});

console.log(sources.find((s) => s.title === "Mem0 Memories")?.providerMetadata?.mem0);

The source has title: "Mem0 Memories" and providerMetadata.mem0 holds memories (array of memory objects) and memoriesText. It is only present when at least one memory was retrieved.

Pattern 2: Standalone Utilities

Use standalone utilities when you want full control over the memory retrieve/store cycle, or you want to use a provider that is already configured separately.

import { openai } from "@ai-sdk/openai";
import { generateText } from "ai";
import { retrieveMemories, addMemories } from "@mem0/vercel-ai-provider";

const prompt = "Recommend a restaurant";

// Retrieve memories -- returns a formatted system prompt string
const memories = await retrieveMemories(prompt, {
  user_id: "alice",
  mem0ApiKey: "m0-xxx",
});

// Generate using any provider with injected memories
const { text } = await generateText({
  model: openai("gpt-5-mini"),
  prompt,
  system: memories,
});

// Optionally store the conversation back
await addMemories(
  [
    { role: "user", content: [{ type: "text", text: prompt }] },
    { role: "assistant", content: [{ type: "text", text }] },
  ],
  { user_id: "alice", mem0ApiKey: "m0-xxx" }
);

Pattern 3: Streaming

Use streamText for streaming responses with memory augmentation:

import { streamText } from "ai";
import { createMem0 } from "@mem0/vercel-ai-provider";

const mem0 = createMem0();
const result = streamText({
  model: mem0("gpt-5-mini", { user_id: "alice" }),
  prompt: "What should I cook for dinner?",
});

for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}

The wrapped model stores the conversation and retrieves memories before streaming begins.

Supported Providers

Select a provider when creating the Mem0 instance:

const mem0 = createMem0({ provider: "anthropic" });
const { text } = await generateText({
  model: mem0("claude-sonnet-4-20250514", { user_id: "alice" }),
  prompt: "Hello!",
});

How It Works Internally

Wrapped model flow

User prompt
  --> processMemories: addMemories (POST /v3/memories/add/, awaited)
  --> processMemories: getMemories (POST /v3/memories/search/)
  --> memories (if any) injected as system message at start of prompt
  --> underlying LLM generates response (doGenerate or doStream)
  --> response returned to caller (doGenerate also attaches a "Mem0 Memories" source)

Standalone flow

User controls each step:
  1. retrieveMemories / getMemories / searchMemories -> fetch memories
  2. inject into system prompt manually
  3. call generateText / streamText with any provider
  4. addMemories -> store new conversation to Mem0

Key Differences Between the 4 Utility Functions

retrieveMemories, getMemories, and searchMemories accept LanguageModelV3Prompt | string as the first argument; addMemories is typed as LanguageModelV3Prompt (a string also works at runtime). All four take optional Mem0ConfigSettings as the second argument.

Common Edge Cases and Tips

• Always provide user_id (or agent_id/app_id/run_id) for consistent memory retrieval. The search endpoint requires at least one entity ID in filters; the provider places these IDs there for you.
• Standalone utilities require explicit API key: pass mem0ApiKey in the config object, or set the MEM0_API_KEY environment variable.
• This uses Vercel AI SDK v6 (LanguageModelV3 / ProviderV3 interfaces, @mem0/vercel-ai-provider 3.x). Provider 2.x targeted AI SDK v5. It is not compatible with AI SDK v4 or earlier.
• processMemories awaits addMemories before searching and calling the LLM, so each wrapped call includes one memory write and one memory search. If either request fails, the error is logged and the LLM call proceeds without memories.
• "google" and "gemini" are both accepted and map to @ai-sdk/google.
• Removed in 3.0.0: org_id, project_id, org_name, project_name, output_format, filter_memories, async_mode, enable_graph, version, api_version. Graph memory is now a Mem0 Platform project setting, not a provider option.
• Default top_k is 10. threshold and rerank are only sent when set (the API default for rerank is false; threshold is a server-side cutoff, not a floor on the returned score).
• Custom host: set host in the config to point to a different Mem0 API endpoint (default: https://api.mem0.ai).

References

Related Mem0 Skills