Mastra

Rekall memory plus durable session state for Mastra, delivered as a memory Processor. The @rekall/mastra package recalls relevant memories into the prompt and captures the assistant's output back into Rekall.

Overview

mem.processor() conforms to Mastra's current Processor interface (processInput / processOutputResult) and does two things: recall-injection (pull relevant Rekall memories into the prompt) and capture (write the assistant's final answer back to Rekall). The processor methods never throw, so memory can never break a run.

Installation

Terminal

npm install @rekall/mastra

Usage

agent.ts

import { Agent } from '@mastra/core/agent';
import { RekallMemory } from '@rekall/mastra';

const mem = new RekallMemory({
  apiUrl: 'https://api.rekall.app',   // or http://localhost:9876 in dev
  apiKey: process.env.REKALL_API_KEY, // rk_… key, or accessToken: 'rat_…'
});

const agent = new Agent({
  name: 'assistant',
  model,
  inputProcessors: [mem.processor()],   // recall → inject as a system message
  outputProcessors: [mem.processor()],  // capture assistant output into memory
});

What This Is (and Isn't)

mem.processor() is an input/output Processor — it is intentionally not a Mastra Memory provider. It does not implement Mastra's first-class Memory abstraction, so it does not participate in Mastra's thread / resource scoping or working-memory.

Rekall owns its session state

Rekall keeps its own durable session state via saveSession / loadSession, which you key however you like (e.g. by thread id). That state is owned by Rekall, not surfaced through Mastra's Memory APIs.

How It Maps

Processor method	Rekall
`processInput`	recall against the latest user message → inject hits as a system message
`processOutputResult`	capture the assistant's final text → `POST /api/v2/memories`

The class also exposes remember / recall / contextFor for direct use. It is duck-typed against Mastra (optional peer dependency, no version coupling).

Session State

Persist durable thread/working state with saveSession and rehydrate it with loadSession, backed by Rekall's execution-memory /external store:

session.ts

await mem.saveSession('thread-1', state);   // durable thread/working state
const prior = await mem.loadSession('thread-1');

Authentication

Pass apiKey (rk_…) or accessToken (rat_…), sent as a Bearer token. For local dev against a header-auth server, pass headers: { 'x-user-id': '…' }.

Next Steps

Browse all framework adapters Understand execution memory and session state Learn about memory types and contexts