Ordalin
Submit a tool

Reviewed tool profile

Hindsight by Vectorize

Open-source memory layer that lets AI agents remember users and learn from past mistakes.

Sources checked

Visit website ↗
Pricing
Freemium. Self-hosting Hindsight is free under the MIT licence with no usage limits. Hindsight Cloud is pay-as-you-go with no monthly or seat fee: retain $10 per million tokens, recall $0.75 per million tokens, reflect $0.05 per call, and storage $0.25 per million tokens a month after the first 30 days. New accounts get $5 of free credit. Enterprise is priced on request. Official source 1 ↗
Available as
Self-hosted with a single Docker command on Postgres, with an MCP server built in. · Python, TypeScript and Go SDKs, MCP and a plain HTTP API. · Hindsight Cloud (managed) and Hindsight Enterprise in a private cloud, your own AWS, Google Cloud or Azure account, or on premises. · Coding-agent integration for Claude Code, Codex CLI, Cursor CLI, opencode, Copilot CLI and more; Hermes and OpenClaw use it as their memory provider. Official source 1 ↗ · Official source 2 ↗ Official source 1 ↗ Official source 1 ↗ Official source 1 ↗
Primary group
Coding & Development

Also consider

See all 6 similar tools ↓

What it does

Vectorize makes Hindsight, an open-source (MIT) memory layer for AI agents. Agents store what happens (retain), look it up later (recall) and turn it into consolidated knowledge (reflect), so they keep per-user context across sessions and learn from failed tool calls and corrections. It works with any LLM, ships its own small decision, rerank and embedding models, and plugs into coding agents, personal agents and your own apps. You can self-host it free, use pay-as-you-go Hindsight Cloud, or buy Hindsight Enterprise.

Official source 1 ↗

Features

  • Retain, recall and reflect. Three API methods store memories, retrieve them, and synthesize patterns from them; extraction, entity linking, retrieval and reranking run behind them. Official source 1 ↗
  • Learns from mistakes. Failed tool calls and user corrections are stored as experiences, so the agent knows what went wrong the next time. Official source 1 ↗
  • Four retrieval strategies in parallel. Dense vector, sparse keyword, graph and time-aware search run together, and results are merged to a token budget rather than a fixed top-K count. Official source 1 ↗
  • Per-user, cross-session memory. Each user gets separate persistent context that survives across sessions; changed facts keep both states with timestamps. Official source 1 ↗ · Official source 2 ↗
  • Project memory for coding agents. One install wires every detected coding agent to a per-repo memory bank built from git history and past sessions, filled automatically in the background. Official source 1 ↗ · Official source 2 ↗
  • Own small models with open weights. A decision model, reranker and embedding model trained for Hindsight; run the open weights from Hugging Face or call them through Hindsight Inference. Official source 1 ↗
  • Published benchmark results. Vectorize reports 94.6% retrieval accuracy on LongMemEval-S and fewer corrections per task for Claude Code on AMB's sdebench. Official source 1 ↗ · Official source 2 ↗

Pricing & free limits

Self-hosted

$0free, MIT licence

Hindsight 1.0 on your own infrastructure with one Docker command; all memory APIs and the MCP server; no usage limits or telemetry; community support via GitHub. Official source 1 ↗

Hindsight Cloud

Pay as you goper operation, prepaid credits

Retain $10 per million tokens, recall $0.75 per million tokens, reflect $0.05 per call, Iris extraction $7.50 per million tokens, mental-model retrieve $0.25 per million tokens and refresh $0.05 per call, storage $0.25 per million tokens a month. No monthly or seat fee; 99.9% uptime SLA and 12×5 email support. Official source 1 ↗

Enterprise

Custominvoiced

Private cloud, your own cloud account or on premises, single-tenant; SSO and enforced MFA, Memory Defense, audit logs and SIEM, SLA up to 99.95%, a shared Slack channel and up to 24×7 support. Official source 1 ↗

Hindsight Models (Inference)

$0.005–$0.02per million input tokens

hs-embed-1 $0.005, hs-rerank-1 $0.015, hs-decide-1-flash $0.02 per million input tokens; output is free, and the open weights are free to run yourself. Official source 1 ↗

Billing details

Hindsight Cloud runs on prepaid credits (1 credit = $1), bought in packages of $10, $25, $50 or $100 or a custom $5–$1,000 through Stripe; Alipay is available in mainland China. Official source 1 ↗

Only successful operations are charged, and tokens are counted with the o200k_base tokenizer. Official source 1 ↗

Memories are free to store for their first 30 days; after that storage is billed, and memories are never deleted. Official source 1 ↗

Auto-recharge can top up the balance when it falls below a threshold. Official source 1 ↗

Free access & limits

New Hindsight Cloud accounts start with $5 of free credit. Official source 1 ↗

The self-hosted version has no usage limits. Official source 1 ↗

Best for & limitations

Give Claude Code, Codex or Cursor a per-repository memory of past decisions from git history and earlier sessions. Official source 1 ↗ · Official source 2 ↗

Use it as the memory provider for personal agents such as Hermes and OpenClaw. Official source 1 ↗

Add per-user memory to your own agent or app through the Python, TypeScript or Go SDK, MCP or HTTP. Official source 1 ↗

  • Cloud stops when credits run out. When the credit balance is exhausted, API calls return HTTP 402 until credits are added. Official source 1 ↗
  • Knowledge-base refreshes can dominate the bill. Knowledge-base pages refresh after each consolidation by default, so that charge grows with how much you write and can exceed every other operation. Official source 1 ↗
  • Enterprise controls are Enterprise-only. SSO and enforced MFA, Memory Defense, audit logs and SIEM come with Enterprise; self-hosting gets community support only. Official source 1 ↗
  • Credit transfers are permanent. Moving credits between organizations cannot be undone. Official source 1 ↗
  • Developer tool, not an end-user app. It is used through SDKs, MCP, HTTP or a coding-agent install, not as a standalone app. Official source 1 ↗
Hindsight by Vectorize product screenshot
Website previewFixed desktop viewport · 1440 × 900

Similar tools

Published tools that share its categories.