Agent Frameworks

Memory for Agent Frameworks

Add Zep agent memory to the frameworks you already use

Zep provides agent memory for agents you build on popular frameworks. Most integrations ship as packages that wire Zep into the framework’s native patterns: they provision users and threads, inject a prompt-ready context block from the user’s Context Graph, persist conversation turns, and can expose graph search as a tool the model calls on demand. Some frameworks — such as Eve and ElevenLabs — use the Zep SDK directly in the framework’s own hooks, proxies, or instructions.

You keep the framework’s orchestration, tools, and runtime. Zep stores and serves memory across sessions — memory of your users, your business, and the work your agents do — so agents stay consistent without you rebuilding retrieval for each stack.

If your framework is not listed below, use the Zep SDKs directly. The same create-user → create-thread → add-messages → get-context loop applies everywhere; see the quick start guide.

Supported frameworks

  • AG2 — An automatic memory loop on AG2’s hook system, with context injection and Zep search tools.
  • AutoGen — Long-term memory classes that plug into AutoGen’s Memory interface, plus tools for search and data ingestion.
  • CrewAI — Storage adapters and tools so CrewAI agents carry context across executions and share a knowledge base.
  • ElevenLabs — Persistent context for ElevenLabs voice agents through a custom LLM proxy.
  • Eve — Channel onMessage + turn-scoped dynamic instructions for turn-relevant graph.search recall, with hooks that persist turns (SDK pattern, no integration package).
  • Google ADK — Real-time message persistence and automatic context injection for Google ADK agents in Python, TypeScript, and Go.
  • LangGraph — Durable, cross-session memory for LangGraph agents via node helpers, a pre-model hook, and a graph-search tool.
  • LiveKit — Persistent memory for LiveKit voice agents, with turn persistence and context injection before each response.
  • Mastra — Processors and tools that add long-term memory to Mastra agents.
  • Microsoft Agent Framework — A context provider that persists turns, injects context on every run, and can register a graph-search tool.
  • NVIDIA NeMo Agent Toolkit — Automatic memory for NVIDIA NeMo Agent Toolkit agents.
  • Pydantic AI — Turn persistence, native prompt context injection, and a model-callable graph-search tool for Pydantic AI.
  • Strands — A Strands MemoryStore for MemoryManager with context injection, batched extraction, and optional graph search.
  • Vercel AI SDK — Middleware, helpers, and tools for long-term memory in Vercel AI SDK (v6) applications.