@agentmem/vercel-ai-provider

Vercel AI SDK integration for AgentMem. Three ways to wire it in — use one or stack them.

npm install @agentmem/vercel-ai-provider ai

Pattern A — middleware (automatic memory injection)

Wrap any @ai-sdk/* model and memories appear in the system prompt on every call. Works the same for generateText, streamText, and generateObject — they all route through the model's doGenerate / doStream hook, which is where the middleware fires.

import { generateText, wrapLanguageModel } from 'ai'
import { anthropic } from '@ai-sdk/anthropic'
import { createAgentMemMiddleware, type LanguageModel } from '@agentmem/vercel-ai-provider'

const cfg = { apiKey: process.env.AGENTMEM_API_KEY!, agentId: 'support-bot' }

const baseModel: LanguageModel = anthropic('claude-opus-4-7')
const model = wrapLanguageModel({
  model:      baseModel,
  middleware: createAgentMemMiddleware(cfg),
})

const { text } = await generateText({ model, prompt: 'What did the customer prefer?' })

Note on LanguageModel. Import the type from this package, not from ai. The ai SDK's LanguageModel alias is a union (string | LanguageModelV3 | LanguageModelV2) for high-level helpers that accept a model id; that union doesn't satisfy wrapLanguageModel, which needs the object form only. This package re-exports the right shape.

Structured output (generateObject) works the same way

Wrap the model exactly as above and use generateObject directly — the middleware injects memory into the prompt before the model sees it, and the structured schema is preserved.

import { generateObject } from 'ai'
import { z } from 'zod'

const { object } = await generateObject({
  model,
  schema: z.object({ verdict: z.enum(['approve', 'reject']), reasoning: z.string() }),
  prompt: 'Should we approve this purchase request?',
})

If search fails, the middleware logs to stderr and lets the call proceed without memory — your model never gets blocked.

Pattern B — tools

import { generateText } from 'ai'
import { anthropic } from '@ai-sdk/anthropic'
import { agentmemMemorize, agentmemRecall } from '@agentmem/vercel-ai-provider'

await generateText({
  model:  anthropic('claude-opus-4-7'),
  tools:  {
    memorize: agentmemMemorize(cfg),
    recall:   agentmemRecall(cfg),
  },
  prompt: '...',
})

The agent decides when to call memorize / recall. Lower per-turn latency than the middleware on simple turns.

Pattern C — helpers (manual control)

Drop-in API replacement for @mem0/vercel-ai-provider — same function names.

import {
  addMemories,
  retrieveMemories,
  searchMemories,
} from '@agentmem/vercel-ai-provider'

await addMemories('User prefers email.', { apiKey, agentId: 'bot' })

const context = await retrieveMemories(userMessage, { apiKey, agentId: 'bot' })
// → "Relevant memories retrieved from AgentMem...\n- User prefers email."
//   Splice into your system prompt manually.

const hits = await searchMemories('what does the user prefer?', { apiKey, agentId: 'bot' })
// → MemoryHit[] for custom rendering

Why this beats @mem0/vercel-ai-provider

| | Mem0 provider | AgentMem provider | |---|---|---| | Wraps every model SDK | Yes (OpenAI, Anthropic, Cohere, Groq, Google) | No | | Works with any @ai-sdk/* provider | Only the 5 it wraps | All of them | | Runtime dependency count | 8+ | 1 (@agentmem/sdk) | | Provides middleware | No | Yes | | Helper API compatibility | — | Same names: addMemories / retrieveMemories / searchMemories |

Migration from Mem0: change the import path, set AGENTMEM_API_KEY instead of MEM0_API_KEY, and pass agentId in the config.

Config reference

| Field | Default | Purpose | |---|---|---| | apiKey | — | Required. | | agentId | — | Required (constructor OR per-call). | | baseUrl | hosted | Self-host override. | | scope | private | Visibility default. | | workflowId | — | Group memories under a workflow. | | topK | 5 | Max memories per recall. | | rerank | false | Re-score with Gemini (+0.5-1.5s). | | minScore | 0 | Drop hits below this score. | | prefix | default | Custom text before the memory block (middleware + retrieveMemories). |

License

Apache-2.0