Vectorize.io

Vectorize.io is a developer tool. Hindsight by Vectorize is an open source, MIT licensed memory layer for AI agents. It stores per-user context across sessions, learns from failed tool calls and corrections through a reflection layer, and plugs into coding agents, personal agents or your own apps via SDKs, MCP or HTTP.

Vectorize's Hindsight is a memory layer for AI agents, built around a simple premise: storing and retrieving facts is not the same as learning. Most memory systems act as a lookup table, so an agent can recall that a tool call failed without ever changing how it behaves next time. Hindsight adds a reflection layer on top of retrieval. Failed tool calls and user corrections are kept as experiences, and the system consolidates raw observations into reusable knowledge, so patterns surface from the data instead of from manual tagging. The product covers per-user memory, where preferences and history stay separated by user, and cross-session persistence, so context survives across weeks of separation. Recall runs as parallel search and the team reports sub-100ms returns. The memory layer is model-agnostic, which means swapping the underlying LLM does not wipe out what the agent has learned. Hindsight is MIT licensed and runs on Postgres; a single Docker command puts it on your own infrastructure. A managed cloud option handles scaling, backups and upgrades on pay-as-you-go usage, and an enterprise tier adds private or bring-your-own-cloud deployment, single-tenant isolation, SSO, audit logs and SLAs. A separate Rust engine, Memlake, targets memory volumes in the billions using object storage. In practice, teams reach Hindsight through three paths. Coding agents get project memory from one npx install that wires Claude Code, Codex CLI, Cursor CLI and other harnesses to a per-repo memory bank built from git history and past sessions. Personal agents such as Hermes and OpenClaw set it up as their memory provider, recalling context before each turn and retaining it afterwards. Custom applications use three primitives (retain, recall and reflect) from Python, TypeScript or Go over MCP or plain HTTP, with extraction, entity linking, four-way retrieval and reranking handled behind the API. The fit is narrow but deep. Hindsight competes with vector database memory add-ons and hosted memory APIs rather than with general agent builders. Teams that want a self-hosted, open source memory engine with published benchmark results and no vendor lock-in to a model provider are the core audience. The team publishes retrieval accuracy comparisons including 94.6% on LongMemEval-S and 92.0% on LoComo, and publishes a comparison only where a competing system has one.