dalao.ai

Hindsight

Agent memory that learns: retain, recall and reflect APIs, MIT-licensed and self-hostableOpen source

About Hindsight

Hindsight is an agent memory system from Vectorize built around three operations: retain uses an LLM to extract facts, entities and relations, recall runs semantic, keyword, graph and temporal searches in parallel, and reflect reasons over a memory bank to answer. The open-source edition is MIT-licensed and self-hostable; Hindsight Cloud bills on token usage with no monthly or seat fee; Enterprise adds private cloud, BYOC and on-premises deployment. As of 2026-10-26 the latest release is v1.1.0, published on GitHub 2026-10-22 00:41 (UTC+8).

EditorialUpdated 7 min read

What problem it solves

The Vectorize About page says every AI agent today forgets everything between conversations. The README says Hindsight aims to make agents that learn over time rather than only recall conversation history, and says it eliminates the shortcomings of RAG and knowledge graphs.

Hindsight does not store raw conversations. It extracts content into typed memories: world facts, the agent's own experiences, observations consolidated from many facts, mental models set up for frequent questions, and knowledge pages the bank maintains about itself. Memories live in isolated banks, one per user, agent or project.

Product shape. The open-source server deploys with Docker, Helm or pip and stores data in Postgres with pgvector. Alongside it are the managed Hindsight Cloud and Hindsight Enterprise. Client SDKs, a REST API and a built-in MCP server all expose the same three operations.

The repository is vectorize-io/hindsight, created 2025-10-30 19:49 (UTC+8). Recent releases: v0.10.3 on 2026-10-08 23:47 (UTC+8), v1.0.0 on 2026-10-13 22:08 (UTC+8) and v1.1.0 on 2026-10-22 00:41 (UTC+8).

The publisher is Vectorize, and LICENSE names Vectorize AI, Inc. as copyright holder. The leadership section of the About page lists Chris Latimer as CEO and Chris Bartholomew as CTO.

Every AI agent today forgets everything between conversations.

Vectorize About page

Core capabilities

Four-way parallel retrieval. Recall runs semantic vector search, BM25 keyword search, graph expansion over entity, temporal and causal links, and time-range filtering at once. Results are fused by rank, reranked by a cross-encoder and cut to a token budget rather than a fixed count.

Observation consolidation. A background worker keeps merging raw facts into observations. Each observation points to the memories behind it with exact quotes, and new evidence updates it while its history is kept.

Reflect. Reflect runs an agentic loop of up to ten rounds over the bank, checking mental models and knowledge pages first, then observations, then raw facts, and it can cite only items it actually retrieved. A bank carries a mission, directives and a disposition of skepticism, literalism and empathy on a 1 to 5 scale, which shape reflect but not recall.

Knowledge pages. A knowledge page is a living document the bank writes about itself, organized in folders and rewritten incrementally after consolidation; hindsight fs mount projects the tree onto disk as plain Markdown.

Integrations. The README cites 25+ LLM providers, including local options such as ollama, lmstudio and llamacpp, and 60+ integrations spanning coding agents such as Claude Code, Codex, Cursor and opencode and frameworks such as LangGraph, LlamaIndex, CrewAI and the OpenAI Agents SDK (as reported by the vendor). v1.1.0 adds an extension that sets the LLM and credentials per call on the server, plus a single decision-model configuration.

Memlake. Released 2026-10-13, Memlake is a storage engine written in Rust that keeps memory in object storage and serves all four search types from one copy. The launch post cites about 90 ms at p99 on a 3M-memory bank and per-bank ingest 15 to 27 times that of Postgres (as reported by the vendor).

Who uses it

As of 2026-10-26 the GitHub API shows 53,704 stars and 6,248 forks. A 2026-09-28 blog post lists 263 contributors, 2,691 merged pull requests and 70 releases (as reported by the vendor).

Company history. The About timeline says the company was founded in January 2024 and closed a $3.6M seed round led by True Ventures in April 2024; a company press release dated 2024-10-08 announces a $3.6M seed round led by True Ventures. The About page dates general availability to September 2024, and the 2024-10-08 press release calls the platform available now.

Vendor statements on adoption (as reported by the vendor): the About page cites 15,000+ developers building with Vectorize and Hindsight; the Memlake post cites more than 30,000 users on Hindsight Cloud, most organizations already running on Memlake; the README says Hindsight is in production at Fortune 500 enterprises.

Benchmark results (as reported by the vendor):

  • LongMemEval-S at 94.6%, against 74.0% for the next best published system.
  • LoCoMo at 92.0%, PersonaMem at 86.6%, LifeBench at 71.5% and BEAM at 10M tokens at 64.1%.
  • On the sdebench coding-agent suite, every run solves 60–61 of the 61 tasks with or without memory; with Hindsight, Claude Code corrections per task drop from 0.85 to 0.36 and cost per task from $0.445 to $0.338.

The README says the benchmark data has been independently reproduced by research collaborators, and the benchmarks page says the research behind it is peer-reviewed and the benchmark code is open source.

How it compares

Against Mem0: Mem0 is Apache-2.0 and describes itself on GitHub as The Memory Layer for AI Agents, the same agent-memory category. Hindsight is MIT-licensed and lists its scores beside the next best system on its benchmarks page.

Against Graphiti: Graphiti is Apache-2.0 and describes itself as Build Real-Time Knowledge Graphs for AI Agents. The Hindsight FAQ describes an event-centric graph: memories are the anchors, entities attach to memories with no direct edges between entities, and change is recorded as new memories while old ones stay.

Against Letta: Letta is Apache-2.0 and describes itself as a Platform for stateful agents. Hindsight is only a memory layer and plugs into other agents and frameworks through SDKs, a REST API, MCP and 60+ integrations.

Against RAG: the FAQ says Hindsight stores structured facts instead of document chunks, consolidates mental models over time and supports time-aware retrieval.

Who it's for / Not for

Good fit:

  • Teams whose agents need to carry user preferences, project conventions and past decisions across sessions.
  • Developers who want memory decoupled from the model, since the server works with 25+ LLM providers and with local models.
  • Organizations that need data to stay in their own environment: the open-source edition self-hosts with no usage limits, and Enterprise offers private cloud, BYOC and on-premises deployment, including air-gapped setups.
  • Teams already on coding agents such as Claude Code, Codex, Cursor or opencode that want memory per repository.

Not a fit:

  • Setups that need fine-grained access control by team, agent or memory without Enterprise: feature request issue 2235 is still open as of 2026-10-26.
  • Environments that forbid any content from reaching a third-party model and cannot host a local one, since retain calls an LLM to extract facts.
  • Self-hosting on Python older than 3.11 or with less than 4GB of RAM.

Getting started

Self-hosting. The README's preferred route is one docker run command with an LLM API key, serving the API on port 8888 and the UI on port 9999; Docker with an external Postgres, Helm and pip install hindsight-api are alternatives. The FAQ lists Python 3.11 or later and 4GB of RAM as the minimum, with 8GB recommended for production.

Clients. Python uses pip install hindsight-client, and a TypeScript client also exists; retain writes to a bank_id, recall searches it, and reflect answers from it. The built-in MCP server lives at a path of the form /mcp/{bank_id}/ and exposes the three operations as tools.

Lower-effort options. hindsight-litellm wraps LLM calls, recalling before each call and retaining afterwards, and npx @vectorize-io/hindsight-coding-agents install wires several coding agents at once.

Managed. A new Hindsight Cloud account comes with free credits for testing before any charges.

Pricing & license

Self-hosted: free and MIT-licensed; the pricing page says no usage limits and no telemetry.

Hindsight Cloud: usage-based billing with no fixed monthly fee, no seat pricing and free credits for new accounts. Rates per operation:

  • Retain: $10.00 per million tokens.
  • Recall: $0.75 per million tokens.
  • Reflect: $0.05 per call.
  • Iris Extract: $7.50 per million tokens.
  • Mental Model retrieve at $0.25 per million tokens, refresh at $0.05 per call.
  • Storage: $0.25 per million tokens stored per month; memories are free to store for their first 30 days and are never deleted.

Cloud includes a 99.9% uptime guarantee, 12×5 email support and a 12-hour response SLA; volume pricing goes through sales.

Enterprise: by quote, with private cloud, BYOC or on-premises, single-tenant deployment, private networking, Memlake, SSO and enforced MFA, Memory Defense, audit logs and SIEM, custom SLAs up to 99.95%, up to 24×7 support with a 30-minute response, and invoiced billing. The Enterprise page states that Vectorize is SOC 2 Type 2 certified.

License: LICENSE is the MIT License, copyright Vectorize AI, Inc. Memlake, SSO, Memory Defense and similar capabilities ship only with Enterprise.

Risks & limits

External models. The README says retain uses an LLM to extract facts, temporal data, entities and relations, and the quick start passes an OpenAI API key; content sent to retain and reflect goes to the chosen model provider unless a local model such as ollama is configured.

Edition tiers. The Memlake post says the open-source edition stays on Postgres; Memlake, SSO, Memory Defense and audit logs come only with Enterprise.

Access control. Issue 2235, opened 2026-06-16 20:58 (UTC+8), requests team, agent and memory-level access control and is still open as of 2026-10-26.

Releases. v1.0.0 shipped 2026-10-13 22:08 (UTC+8) and v1.1.0 shipped 2026-10-22 00:41 (UTC+8); the Memlake post says 2.0 is already in development.

Differing statements in vendor materials:

  • Company name: LICENSE reads Vectorize AI, Inc.; the site footer reads Vectorize, Inc.
  • SDK languages: one FAQ answer lists SDKs for Python, TypeScript, Go and Rust; the FAQ client list names Python, TypeScript, Go, CLI and HTTP.

Our verdict

Hindsight lets agents work with memory through three operations, retain, recall and reflect; recall runs four retrieval strategies in parallel and cuts results to a token budget; and the open-source edition is MIT-licensed, self-hostable and can use local models such as ollama.

Its limits: retain and reflect both call an LLM, so content goes to the chosen model provider unless a local model is used; the fine-grained access control request, issue 2235, is still open as of 2026-10-26; v1.0.0 shipped on 2026-10-13; the benchmark results are vendor-reported; and Memlake, SSO and Memory Defense come only with Enterprise.

Sources

Official

  1. Overview | Hindsighthindsight.vectorize.io · Accessed Oct 26, 2026
  2. Frequently Asked Questions | Hindsighthindsight.vectorize.io · Accessed Oct 26, 2026
  3. Introducing Hindsight Memlake: serving 100M Memories at <100ms p99 on object storagehindsight.vectorize.io · Accessed Oct 26, 2026
  4. 40,000 Stars, and the Number We Actually Watchhindsight.vectorize.io · Accessed Oct 26, 2026
  5. Pricing - Hindsight Agent Memory | Vectorizevectorize.io · Accessed Oct 26, 2026
  6. About | Vectorizevectorize.io · Accessed Oct 26, 2026
  7. Hindsight Enterprise | Vectorizevectorize.io · Accessed Oct 26, 2026
  8. Hindsight Benchmarks - Agent Memory Accuracy | Vectorizevectorize.io · Accessed Oct 26, 2026

Updates & reading