MemoryRouter
CrewAI memory

CrewAI memory: one API, and the storage question

CrewAI did the useful consolidation: one Memory class instead of four separate systems, with LLM-assisted analysis on save and scoring that blends similarity, recency, and importance at recall.

CrewAI. Sources checked 2026-09-25.

How CrewAI handles memory today

The current docs describe a unified Memory system that replaced separate short-term, long-term, entity, and external memory. Saving runs analysis that infers scope, categories, importance, and metadata, and recall uses adaptive-depth composite scoring rather than pure vector similarity.

Storage is your side of the contract. For a crew run on your laptop that is fine. For a product with many end users, the memory layer becomes part of your production surface.

What CrewAI forgets

The API is solved. The lifecycle is not.

The fix with MemoryRouter

Two integration points. Point the crew’s model provider at the MemoryRouter OpenAI-compatible endpoint so recall happens inside the request path, and use the API for explicit memory operations your agents need.

  1. 1

    Route model calls through MemoryRouter

    Environment

    export OPENAI_API_BASE=https://api.memoryrouter.ai/v1
    export OPENAI_API_KEY=mk_your_memory_key
  2. 2

    Scope memory per end user

    Create one vault per user and swap keys per request, so user A never retrieves user B’s context. Vaults are provisioned through the dashboard or the API.

  3. 3

    Keep crew memory for crew state

    Leave the built-in Memory class in place for run-level state, and let the vault hold the durable record you want to keep across sessions and products.

Start a MemoryRouter vault

Per-user vaults, keys, and an OpenAI-compatible endpoint. 14-day free trial.

Related

CrewAI memory FAQ

Sources

Sources checked 2026-09-25. Tool behavior changes; check the vendor docs before relying on a detail.

Start free