Hindsight Memory Adapter

Hindsight Memory Adapter

Hosted recall/save adapter backed by Hindsight. Hindsight owns extraction and ranking server-side, buckets memory into per-conversation "banks" ({tenantId|_}__{user}__{threadId}), and — uniquely — exposes LLM tools through recall so the model can retain/recall/reflect directly.

Setup

import { memoryMiddleware } from '@tanstack/ai-memory'
import { hindsight } from '@tanstack/ai-memory/hindsight'

// Build per request from the server-validated session — never from req.body.
function memoryFor(session: { userId: string; threadId: string }) {
  const memory = hindsight({ user: session.userId }) // baseUrl defaults to HINDSIGHT_URL

  return memoryMiddleware({
    adapter: memory,
    scope: { threadId: session.threadId, userId: session.userId },
  })
}

@vectorize-io/hindsight-client is an optional peer dependency, loaded lazily on first use — install it where you use hindsight().

Options

• user — durable user id for the bank key (falls back to scope.userId).
• baseUrl — Hindsight server URL (default HINDSIGHT_URL or http://localhost:8888).
• budget — recall budget: 'low' | 'mid' | 'high' (default 'mid').
• onToolRetain / onToolRecall — callbacks fired when the model uses the memory tools.

Scope fields: bank id is {tenantId|_}__{user}__{threadId}. namespace is ignored.

Tools

recall returns hindsight_retain, hindsight_recall, and hindsight_reflect in its tools plus a toolGuidance block. memoryMiddleware merges them into the run so the