id: hindsight_pinned
framework: hindsight
mode: white_box
status: implemented_store_harness
description: Hindsight memory target using the official client, retain, recall, list_memories, and document deletion.
source_url: https://hindsight.vectorize.io/
source:
  type: python_package_and_service
  repo: vectorize-io/hindsight
  package: hindsight-client
  version: 0.8.4
runtime:
  type: docker_or_external_service
  base_url: http://localhost:8888
  required_env:
    - OPENAI_API_KEY
model:
  provider: openai
  model: gpt-4o-mini
  temperature: 0
adapter:
  module: memorybench.adapters.hindsight
  class: HindsightAdapter
capabilities:
  inspect_memory: true
  delete_memory: document_delete
  retrieval_log: true
  multi_user: true
  multi_tenant: true
notes:
  - The adapter expects a Hindsight API server on HINDSIGHT_BASE_URL or http://localhost:8888, creates one temporary bank per benchmark subject, writes trusted facts with retain(), retrieves with recall(), inspects native list_memories(), and deletes stale entries by document_id.
  - The official client package is hindsight-client==0.8.4. Do not use the unrelated/broken PyPI package named hindsight==0.1.7.
  - The local score was produced with ghcr.io/vectorize-io/hindsight:0.8.4-slim on Colima, OpenAI gpt-4o-mini, OpenAI text-embedding-3-small embeddings, and RRF reranking.