MITfreeRuns locallyAI Chat & Agents

Hindsight: MIT Agent Memory

MIT agent memory that learns — retain, recall, reflect, and banks. Not a chat model.

Official Hindsight memory-types overview figure from the vectorize-io/hindsight README — Official project image. Source · License: MIT.

License: MIT
Stars: 56,654 (2026-10-31)
Platforms: Linux, macOS, Windows, Docker, Kubernetes, Python, Node.js, Go
First seen: 2025-10-30

TL;DR. Hindsight is MIT agent memory that learns — retain, recall, reflect, and banks. Not a chat model. License: MIT (verified). Platforms: Linux, macOS, Windows, Docker, Kubernetes, Python, Node.js, Go. Price: free. Runs locally: yes. It is an alternative to Mem0, Claude Code, Cursor.

Name	Hindsight
Language	Python
License	MIT
Pricing	free
Platforms	Linux, macOS, Windows, Docker, Kubernetes, Python, Node.js, Go
Runs locally	Yes
Stats	56,654 stars · 6,813 forks (as of 2026-10-31)
First seen	2025-10-30
Repository	https://github.com/vectorize-io/hindsight

Hindsight is an open-source agent memory system from Vectorize. Official hosts we opened on 2026-10-31: vectorize-io/hindsight (56,654 stars, 6,813 forks, created 2025-10-30T11:49:48Z, last push 2026-10-30T23:40:43Z, default branch main), LICENSE (MIT), tag v1.1.0 (published 2026-10-21T16:41:42Z), docs at hindsight.vectorize.io, and the paper listing arXiv 2512.12818. GitHub description: “Hindsight: Agent Memory That Learns.” Topics: agentic-ai, agents, ai-memory, memory. Language: Python.

It hangs off Mem0 alternatives as a memory layer, and off Claude Code alternatives and Cursor alternatives because the README names those products among coding-agent / provider integrations. That mapping is editorial. We have no sourced quality comparison against Mem0, Claude Code, or Cursor.

The README (opened 2026-10-31) says most agent memory systems focus on recalling conversation history, while Hindsight is focused on making agents that learn, not just remember. Three operations are named: retain, recall, reflect. Memories live in banks. Those sentences stay upstream.

Key facts (2026-10-31)

Field	Value	Source
Stars / forks	56,654 / 6,813	GitHub API
Created	2025-10-30T11:49:48Z	same
Last push	2026-10-30T23:40:43Z	same
Latest release	v1.1.0 (2026-10-21T16:41:42Z)	GitHub Releases API
License	MIT	LICENSE + GitHub API
Language	Python	GitHub API
Job	Agent memory (retain / recall / reflect, banks)	README
Homepage	hindsight.vectorize.io	GitHub `homepage`

README memory-type labels, copied only: world facts, experiences, observations, mental models. Recall is described as four parallel strategies (semantic, keyword/BM25, graph, temporal) merged with reciprocal rank fusion and a cross-encoder reranker. Those are vendor architecture labels, not a forkalt test.

License

LICENSE is MIT, copyright 2025 Vectorize AI, Inc. GitHub license.spdx_id MIT. Hosted Hindsight Cloud is a different product. Not legal advice.

Install (docs, not tested)

README Docker one-liner (requires an LLM API key in the environment the docs name):

docker run -it --pull always --name hindsight --restart unless-stopped \
  -p 8888:8888 -p 9999:9999 \
  -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
  -v hindsight-data:/home/hindsight/.pg0 \
  ghcr.io/vectorize-io/hindsight:latest

Copied defaults: API http://localhost:8888, UI http://localhost:9999. Bare metal: pip install hindsight-api then hindsight-api. Helm: helm install hindsight oci://ghcr.io/vectorize-io/charts/hindsight. Clients: pip install hindsight-client -U, npm install @vectorize-io/hindsight-client. Embedded (no server): pip install hindsight-all -U. The README also names Claude Code, Cursor, and GitHub Copilot as existing-subscription providers that need no API key — those labels are upstream. forkalt has not launched the stack.

What we did not invent

LongMemEval “state-of-the-art” copy and the README benchmark figure stay vendor. We did not read a numeric score off the chart.
We do not rank Hindsight against Mem0, Hy-Memory, or RAGFlow.

Mem0 alternatives · Claude Code alternatives · Cursor alternatives
Hy-Memory (30B-A3B writer model, not a memory service)
RAGFlow on Notion AI alternatives (document RAG, different job)
DeepSeek Harness · Orca (agent runtimes, not memory stores)
AI Chat & Agents

FAQ

Is Hindsight open source?

Yes on the file we opened. LICENSE is MIT (copyright 2025 Vectorize AI, Inc.). GitHub license field MIT. 56,654 stars / 6,813 forks, last push 2026-10-30T23:40:43Z, latest release v1.1.0 (2026-10-21). Hosted Hindsight Cloud is a different artifact. Not legal advice.

Is this a Mem0, Claude Code, or Cursor alternative?

Product-class mapping only. The README describes an agent memory system (retain / recall / reflect, banks) and names Claude Code and Cursor among coding-agent and LLM-provider integrations. Mem0 is the memory-layer search class this page hangs off. We have no sourced quality comparison against Mem0, Claude Code, or Cursor.

How do I run it?

README (not run here): Docker image ghcr.io/vectorize-io/hindsight:latest on ports 8888 (API) and 9999 (UI); or `pip install hindsight-api` then `hindsight-api`; or Helm chart oci://ghcr.io/vectorize-io/charts/hindsight. Clients: `pip install hindsight-client`, npm `@vectorize-io/hindsight-client`. Embedded path: `pip install hindsight-all`. forkalt has not launched it.

Does it replace RAGFlow or Hy-Memory?

Different jobs. RAGFlow is a document RAG/agent engine. Hy-Memory is a 30B-A3B memory-writer model whose card names Mem0 as one harness. Hindsight is a memory service (banks + retain/recall/reflect). No quality rank.

Sources & verification

Last verified 2026-10-31 · Last updated 2026-10-31· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.