[CahidArda]

Stack: #100 → #102 (TanStack stores) → this PR (2/3) → eve 1.0 RC (3/3). This PR targets #102's branch and adds one commit.

What

The AI SDK docs list memory providers: Letta, Mem0, Supermemory, Hindsight, MongoDB and Redis Agent Memory, in content/docs/03-agents/06-memory.mdx in vercel/ai. They share one shape:

• a factory created once
• per-request tools
• an instructions string
• "automatic memory", which most of them do by wrapping the model with wrapLanguageModel

@upstash/agentkit-ai-sdk now offers that shape as createAgentMemory(), built on the existing AgentMemory (Redis Search $smart, so no vector DB):

• memory.tools({ userId }): the existing recall_memory / save_memory tools.
• memory.prepareStep({ userId, query?, topK?, minScore?, onError? }): automatic recall without wrapping the model, because CLAUDE.md forbids model wrappers.
  • On the first step, it appends a <memories> block ("treat as data") to the instructions. ai 7's prepareStep override carries forward to later steps.
  • userId can be a function of the step options, e.g. runtimeContext on a shared ToolLoopAgent.
  • A failed recall never fails the call; the error goes to onError.
• memory.recallContext(), memory.instructions (names your tools), and memory.store.
• createMemoryTools keeps its API and now shares buildMemoryTools with the provider.

The ai-sdk-demo route uses the provider. The README has a "Memory provider" section, and CLAUDE.md is updated.

Other AI SDK v7 extension points checked, not built

• Automatic transcript extraction: needs an LLM call or a model wrapper.
• Message persistence: createChatHistory already covers it.
• Resumable streams: the resumable-stream package already does this.
• Response caching: it's a model middleware, and the model cache was removed on purpose.
• registerTelemetry usage meter / token-based rate limiting: a possible small follow-up.
• Realtime usage (RealtimeModelV4Usage): experimental.
• experimental_queue: client-only.
• Tools registry: docs only. A vercel/ai docs PR is being prepared separately.

Tests

• agent-memory.test.ts (7, live Redis + MockLanguageModelV4) covers:
  • the tools and recallContext
  • a two-step run where the block is injected once and carried forward
  • ToolLoopAgent with runtimeContext
  • an empty user
  • a failed recall going to onError
• ai-sdk suites: agent-memory 7/7, memory 2/2, search-tools 3/3, telemetry 5/5, tools 3/3, one file per fresh free DB.
• Re-checked after stacking on feat(tanstack-ai): sandbox snapshot and MCP stores for TanStack AI 1.x [stack 1/3] #102: install, lint, typecheck, build, sync-version, and agent-memory + memory tests are green.

Changeset

@upstash/agentkit-ai-sdk: minor. It's in the linked group.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CMK4etJMDDBFkjEyWDVCdy

[linear-code]

DX-3285