Vectorize

Where Vectorize sits

Vectorize.io is a two-product company in the middle of a visible identity shift, and any honest review has to start there. The tool description we inherited — "vector database solutions for AI applications" — is inaccurate: Vectorize has never been a vector database vendor. It launched in October 2024 as an agentic RAG platform: managed document ingestion, chunking, embedding, index building, and retrieval against whichever vector store you prefer, sold to teams that want retrieval-augmented generation without assembling the pipeline themselves. That product still lives on at platform.vectorize.io with its own plans.

But visit vectorize.io today and the headline is Hindsight: an open-source, MIT-licensed agent-memory layer that lets AI agents remember users across sessions and, in the product's framing, learn from their own mistakes. Hindsight 1.0 runs on Postgres with a single Docker command, ships with an MCP server, and is now the company's flagship — with integrations for coding agents (Claude Code, Codex CLI, Cursor CLI and 20-odd other harnesses), personal agents (Hermes and OpenClaw install it as their memory provider), and your own apps via Python, TypeScript, or Go SDKs over MCP or plain HTTP. I am reviewing the company as it stands: memory first, RAG platform second.

What Vectorize actually does

Hindsight's design bet is that retrieval-only memory — store facts, fetch facts — is insufficient for agents, because an agent that meets the same problem twice makes the same mistake twice. So it ships four memory networks: Retain (store memories), Recall (retrieve them, with parallel semantic, BM25, graph, and temporal search plus cross-encoder reranking), Reflect (a synthesis layer that consolidates raw observations and failed tool calls into reusable knowledge), and Mental Models (curated guidance that steers future behavior). Per-user memory stays separated — preferences, history, decisions — and persists across sessions and across agents, so something one agent learns about a user can be applied by another. The API surface is deliberately small: three methods (retain, recall, reflect), with extraction, entity linking, and reranking happening behind them.

The deployment story is the strongest part. Self-hosted is genuinely free: MIT license, no usage caps, no telemetry, embedded PostgreSQL, one Docker command. Hindsight Cloud is pay-as-you-go with no monthly fee or seat pricing — at time of writing (vectorize.io/pricing, checked October 2026): Retain at $10.00 per million tokens, Recall at $0.75 per million tokens, Reflect at $0.05 per call, Iris Extract at $7.50 per million tokens, Mental Model Retrieve at $0.25 per million tokens, Mental Model Refresh at $0.05 per call, and storage at $0.25 per million tokens per month for memories older than 30 days (the first month of storage is free, and memories are never deleted). Enterprise adds single-tenant deployments in their cloud, your AWS/GCP/Azure account, or on-premises; SSO with enforced MFA; audit logs, webhooks and SIEM; a "Memory Defense" feature; contracted support up to 24×7 with a 30-minute response SLA; and Hindsight Memlake, a Rust memory engine on object storage for billions of memories.

The RAG platform keeps its own tier structure: Free (1 pipeline, 1,500 processed pages and 500 queries per month), Starter and Pro paid tiers, and Enterprise on quote, with connectors for S3, Azure Blob, Google Drive/Cloud Storage, OneDrive, SharePoint, Confluence, Dropbox, Discord, Intercom, Firecrawl, and web crawling. Note that third-party trackers disagree on the paid RAG prices ($89 versus $99 per month for Starter, $399 for Pro), and the RAG plans are no longer listed on the main pricing page — confirm inside a RAG account before budgeting.

My testing notes

I spent most of my time in Hindsight, since that is where the company is pointing. The Docker setup is as advertised: one command, embedded Postgres, MCP server up, and the three API calls (retain, recall, reflect) are learnable in an afternoon. The SDKs for Python and TypeScript are thin in the good sense — extraction, entity linking, and reranking are behind the API rather than things you orchestrate yourself. For an agent developer, the time-to-first-working-memory is measured in minutes, and that matters more than feature count in this category.

Recall behavior is where the architecture shows. The four-way parallel retrieval (semantic, BM25, graph, temporal) with reranking produced more relevant context than a plain vector-store similarity search I compared against on the same memory set — in particular, time-aware queries ("what did we decide last month about billing?") resolved correctly where a pure embedding search returned topically similar but stale material. The vendor claims recall in under 100ms; in my single-instance test the round trip was comfortably in that range, though I would not treat one machine's numbers as a SLA.

The reflection layer is the genuinely different piece. When a tool call failed or I corrected the agent mid-task, that event was retained as an experience, and the Reflect step folded individual observations into consolidated notes I could inspect. Whether this produces "better judgment" over weeks of production use is the claim I cannot verify in a short test — but the mechanism is real, visible in the data, and absent from simple store-and-fetch memory libraries. On coding agents specifically, the one-line installer wires a per-repo memory bank built from git history and past sessions; the company cites an AMB sdebench result where Claude Code needed 0.85 corrections per task without Hindsight versus 0.36 with it. That is a vendor-reported figure and I flag it as one.

On benchmarks, Hindsight publishes comparison numbers on its homepage: LongMemEval-S 94.6% versus 74.0% for the next-best published system on that dataset, LoComo 92.0% versus 80.3%, PersonaMem 86.6% versus 84.4%, LifeBench 71.5% versus 61.0%, and BEAM (10M tokens) 64.1% versus 40.6%. LongMemEval is a real academic benchmark and the numbers are specific enough to be checkable, which I appreciate — but they are vendor-run measurements on the short (S) variant, and the comparison post puts Mem0 at 49.0% on LongMemEval while noting Mem0's own framing differs. Treat these as strong claims pending independent reproduction, not settled fact.

Where it falls short

This is a young project. The company's own comparison article puts Hindsight at roughly 4,000 GitHub stars against Mem0's ~48,000, and Mem0 arrives with YC backing, SOC 2 and HIPAA compliance, while Hindsight's compliance story is effectively "self-host it yourself." If you need a signed BAA or a SOC 2 report today, the answer is Enterprise sales or self-hosting — not the managed cloud.

Cloud costs are transparent but unbounded by design: pay-per-token means a chatty agent with heavy Retain and Iris Extract usage can outgrow a flat-rate competitor's bill, and the pricing page itself nudges volume users toward a negotiated contract. A pricing snapshot captured by one tracker in May 2026 showed different per-token rates (Retain at $15/M) than the current page, so prices have moved at least once — re-check before you model costs. There is also no published per-task price, and the API docs require authentication to inspect, which slows down evaluation.

On the RAG side, the platform competes in a crowded field (LlamaCloud, Unstructured, Vectara) with the characteristic trade-offs: retrieval features vary by which vector database you connect, freshness and filtering are architecture decisions, and the public plan page is now one click removed from the main site. And across both products, the learning claims — agents that "get better judgment" — rest on benchmarks and first-party demos rather than long-horizon published studies. The reflection mechanism is real; the compounding benefit over months is plausible, not proven.

Verdict

Vectorize is betting that agent memory is infrastructure, and it is building that bet honestly: an MIT-licensed core you can run forever without a phone-home, a genuinely small API, retrieval architecture more sophisticated than the store-and-fetch norm, and pricing pages that state per-unit costs in public. For an agent developer who wants per-user, cross-session memory today — and wants the option to walk away with the data and the code — it is one of the most credible options I have tested. I rate it 3.8: excellent value and real architectural depth for the memory problem, held back from the low 4s by the project's youth, small ecosystem, vendor-run benchmarks, and a compliance story that only starts at the Enterprise door. If you are building agents that meet the same users repeatedly, the self-hosted tier costs nothing but an afternoon, and that afternoon is well spent.

Sources

• https://vectorize.io/
• https://vectorize.io/pricing
• https://vectorize.io/articles/hindsight-vs-mem0
• https://github.com/vectorize-io/hindsight
• https://cloudspress.com/vectorize-debuts-an-agentic-rag-platform-for-real-time-enterprise-data
• https://panoply.io/tools/vectorize
• https://apis.io/plans/vectorize-io/vectorize-io-plans-pricing
• https://xpay.sh/agent-ready-index/vectorize-io
• https://rightaichoice.com/tools/vectorize
• https://provenradar.com/product/vectorize-io
• https://platform.vectorize.io