Hindsight: Vectorize's Open-Source Agent Memory, Explained
Hindsight is an open-source memory system for AI agents, made by Vectorize. Its pitch is that agents should learn, not just remember: rather than storing snippets of conversation and searching them later, it sorts what an agent knows into facts about the world, the agent’s own experiences, consolidated beliefs backed by evidence, and “mental models,” standing answers it keeps rewriting as it learns more. It runs from a single Docker container under the MIT license, with a usage-billed Hindsight Cloud for teams that don’t want to host it. It has become one of the most popular projects in the category, with about 60,700 GitHub stars in its first year.
The quick answer: Hindsight is a strong pick for developers who want a capable, self-hostable memory layer with an unusually thoughtful design, published research behind it and dozens of ready-made integrations, from Claude Code to LangGraph. Its benchmark scores are among the highest reported, though, as everywhere in this category, they come from the company and its research partners. A disclosure: we make Carly, an AI assistant with its own memory, so we compete on the memory side. Everything below comes from Hindsight’s README, documentation, pricing, research paper, Vectorize’s about page and press coverage.
Hindsight at a glance
Checked against Hindsight’s GitHub repository, documentation and Vectorize’s pricing and about pages on November 12, 2026.
| Hindsight | |
|---|---|
| What it is | An open-source memory system that helps AI agents learn over time |
| Core operations | retain, recall and reflect |
| Memory structure | World facts, experiences, observations and mental models, kept in isolated banks |
| Open source | MIT license, about 60,700 GitHub stars; version 1.0 shipped October 2026 |
| Works with | 60+ integrations, including Claude Code, Codex, Cursor, LangGraph, CrewAI, n8n and ChatGPT; built-in MCP server |
| Hosting | Self-hosted (free), Hindsight Cloud (usage-based) or Enterprise in your cloud |
| Company | Vectorize, founded January 2024 by Chris Latimer (CEO) and Chris Bartholomew (CTO) |
What Hindsight does
Hindsight’s README states the idea plainly: “Most agent memory systems focus on recalling conversation history. Hindsight is focused on making agents that learn, not just remember.”
The research behind it, a December 2025 paper titled “Hindsight is 20/20: Building Agent Memory that Retains, Recalls, and Reflects,” argues that typical memory layers “blur the line between evidence and inference.” If a user once said something offhand, and the agent later formed an opinion from it, most systems store both the same way. Hindsight keeps them apart, so an agent knows what it observed, what it concluded and how confident it is.
How the memory is organized
Hindsight stores memories in banks, each an isolated memory for one user, agent or project. Within a bank, memories take four forms:
- World facts, objective information (“The stove gets hot”).
- Experiences, the agent’s own history, written from its point of view (“I touched the stove and it really hurt”).
- Observations, beliefs consolidated from many memories in the background. Each keeps its supporting evidence with exact quotes and a count of how many memories back it, and new information strengthens, weakens or extends a belief instead of silently replacing it.
- Mental models, standing answers to questions you define, such as “What are this user’s preferences?” Hindsight writes the answer, stores it and rewrites it as the bank learns. Reading one is a plain database lookup with no model call, so an agent can start a session already knowing the settled facts. Pages present these as wiki-style documents that can be exported as markdown files.
Banks also carry “disposition” settings, such as skepticism, literalism and empathy, that shape how the agent reasons over its memories.
Retain, recall and reflect
Three operations do the work:
- Retain adds information. A model extracts facts, dates, entities and relationships, then normalizes them into a structure that supports later search.
- Recall retrieves memories by running four searches at once: meaning-based vector search, exact keyword matching, links between entities and events, and time ranges, so a question like “What happened in June?” works. The results are merged and reranked to fit the context window.
- Reflect goes deeper, reasoning across many memories to answer questions that need thought rather than lookup. Hindsight’s examples include a project-management agent reflecting on which risks need attention, and a support agent spotting questions the documentation doesn’t answer.
Hindsight also offers Memory Defense, an optional per-bank setting that scans incoming memories for secrets and personal data against 45 patterns and redacts or blocks them before storage. It’s multilingual by default and keeps names in their original script.
Hindsight Models
In November 2026, Vectorize released three small models built for memory work: a decision model that handles the many small judgments inside the system, like whether a fact is new or which memories matter for a question; a reranker; and an embedding model. They’re open weights on Hugging Face under Apache 2.0 or MIT licenses, or hosted on Hindsight Inference. With the decision model turned on, Hindsight sends far fewer of those judgments to an expensive general model.
Integrations
Hindsight lists more than 60 integrations, most needing no code changes:
- Coding agents: Claude Code, Codex, Cursor, GitHub Copilot, Cline, Aider, Zed and others. One package sets up per-repository memory built automatically from git history and past sessions.
- Agent frameworks: LangGraph and LangChain, LlamaIndex, CrewAI, Pydantic AI, the OpenAI Agents SDK, Google ADK, the Vercel AI SDK and more.
- No-code tools: n8n, Zapier, Dify and Flowise.
- Apps: ChatGPT, Perplexity, Obsidian, and voice platforms Pipecat and Vapi.
The quickest route is an “LLM wrapper”: swap an OpenAI or Anthropic client for a wrapped one, and memories are stored and recalled on every call. Every server also has a built-in MCP endpoint for each bank, and there are Python, TypeScript and Go clients, a CLI and a REST API. Hindsight works with more than 25 model providers, including local ones like Ollama.
Benchmarks
Hindsight’s paper reports 91.4% accuracy on LongMemEval and up to 89.61% on LoCoMo with a large model, and shows a 20-billion-parameter open model jumping from 39% to 83.6% compared with feeding it the full conversation. Vectorize’s about page now cites 94.6% on LongMemEval’s S split. VentureBeat covered the launch, noting the biggest gains came on questions spanning multiple sessions and on reasoning about time.
The README says the results were “independently reproduced by research collaborators” at Virginia Tech’s Sanghani Center and The Washington Post. Readers should know that Naren Ramakrishnan, the Virginia Tech professor who directs the Sanghani Center, is a co-author of the paper, so this is collaborative rather than fully independent validation. Redis, publishing its own memory benchmark in 2026, noted that it couldn’t reproduce some vendors’ higher published scores under its protocol; that applies across the industry, and it’s worth testing any memory system on your own data.
Pricing
Vectorize’s pricing page offers three ways to run Hindsight:
| Option | Price | What’s included |
|---|---|---|
| Self-hosted | Free | The full system under MIT license, one Docker command, built-in MCP server, community support |
| Hindsight Cloud | Pay as you go on tokens | Managed infrastructure, automatic scaling, daily backups, dashboard, team collaboration, 99.9% uptime |
| Enterprise | Contact sales | Private cloud, your own AWS, Google Cloud or Azure, or on premises; SSO, audit logs, custom SLAs and 24/7 support |
Cloud has no monthly or per-seat fee, and new accounts start with free credit.
The company
Vectorize was founded in January 2024 by Chris Latimer, its CEO, and Chris Bartholomew, its CTO, who designed Hindsight’s architecture. According to its about page, it raised a $3.6 million seed round led by True Ventures in April 2024 and launched a data platform for AI applications that September, before releasing Hindsight as open source in 2025. Vectorize says more than 15,000 developers build with its tools and that Hindsight runs in production at Fortune 500 companies.
Hindsight’s growth stands out: its GitHub repository was created at the end of October 2025 and has about 60,700 stars a year later. We found little independent user feedback beyond developer forums, where early comments were positive.
How Hindsight compares
Hindsight competes with memory layers like Supermemory, Honcho, Zep, Cognee, MemoryLake, Redis Agent Memory and Mem0, and with libraries like LangMem. Like Zep and Cognee, it builds structure rather than just storing text, but its focus on separating evidence from belief and keeping self-updating mental models is distinctive. Like Cognee, its full engine is open source, under an even more permissive MIT license. Honcho is its closest match on reasoning about users, and Redis Agent Memory is the simpler, infrastructure-first alternative.
Hindsight is for developers building agents that learn. If you want an assistant that already does, Carly has memory built in: it saves named memory documents that it reads and searches in later conversations, searches a knowledge base of files you upload by meaning, and keeps your to-dos, data tables and CRM in one workspace. Nothing to deploy or tune, it works inside ChatGPT or Claude or by text, and it also labels and sorts email across all your accounts, handles advanced scheduling, and makes and answers calls and texts.
Who Hindsight suits
Hindsight is a good fit if:
- You want an open-source memory system you can self-host with one command.
- Your agents need to learn from experience and explain what they believe and why.
- You use Claude Code, Codex or Cursor and want project memory across sessions.
- You build with LangGraph, CrewAI or another major framework and want a ready integration.
Look elsewhere if:
- Your workflow is simple; Hindsight’s own README says it “may be overkill” for basic n8n-style automations.
- You want a finished assistant rather than infrastructure.
- You need benchmark results validated by a party with no ties to the vendor.
FAQ
What is Hindsight?
Hindsight is an open-source agent memory system from Vectorize. It organizes memory into world facts, experiences, observations and mental models, and gives agents retain, recall and reflect operations so they can learn over time.
Is Hindsight free?
Yes. Hindsight is MIT-licensed and free to self-host. Hindsight Cloud is billed on token usage with no monthly fee, and Enterprise is priced through sales.
Does Hindsight work with Claude Code?
Yes. Hindsight has integrations for Claude Code, Codex, Cursor and other coding agents, plus a built-in MCP server.
How accurate is Hindsight?
Hindsight’s paper reports 91.4% on LongMemEval and up to 89.61% on LoCoMo, and Vectorize now cites 94.6% on LongMemEval-S. These results come from Vectorize and its research collaborators.
Who makes Hindsight?
Vectorize, founded in 2024 by Chris Latimer and Chris Bartholomew, makes Hindsight. The company raised a $3.6 million seed round led by True Ventures.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
