Hindsight

Active
GitHub Python MIT

Description

Hindsight is an agent memory system that learns autonomously, supporting memory retention, recall, and reflection to give AI agents persistent experiential memory.

Key Features

  • Agent memory system focused on learning rather than just remembering, with retain/recall/reflect operations
  • State-of-the-art performance on LongMemEval benchmark, independently reproduced by Virginia Tech and Washington Post
  • LLM Wrapper for 2-line integration swapping your current LLM client with Hindsight memory-enabled client
  • Multi-provider support including OpenAI, Anthropic, Gemini, Groq, Ollama, LM Studio, and MiniMax
  • Embedded Python mode running without a server for lightweight integration
  • Per-user memory isolation with custom metadata for personalized conversational AI agents

Use Cases

💡 Personalizing AI chatbots with per-user memory that persists across sessions
💡 Building AI employees that learn from feedback and handle open-ended tasks autonomously
💡 Adding long-term experiential memory to coding agents for improved context retention
💡 Enterprise conversational AI requiring accurate recall of past interactions and preferences
💡 Autonomous agents that change behavior based on accumulated experience over time

Strengths & Limitations

Strengths

  • Actively maintained, recent updates
  • High community interest (22.2k stars)
  • Permissive open-source license (MIT)
  • Responsive to issues, low backlog

Quick Start

Run `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 $HOME/.hindsight-docker:/home/hindsight/.pg0 ghcr.io/vectorize-io/hindsight:latest`, then install the client with `pip install hindsight-client` and use `client.retain()`, `client.recall()`, and `client.reflect()` to manage memories.

Related Projects

MemAgent

1.1k · Python
Stale B

A MemAgent framework that can extrapolate to 3.5M context tokens, along with a training framework for RL training of any agent workflow.

memoryagentrag +2
  • · Ultra-Long Context Processing: Extrapolate from 8K training context to 3.5M tokens with performance loss under 5%
  • · Reinforcement Learning Driven: Trained with RLVR (Reinforcement Learning from Verifiable Rewards), extends DAPO algorithm for end-to-end multi-turn conversation optimization
  • · Linear Time Complexity: Breaks through computational bottlenecks in long-text processing with linear resource scaling

OpenMemory

4.5k · TypeScript
Active A+

Local persistent memory store for LLM applications including Claude Desktop, GitHub Copilot, Codex, and more. Provides durable context memory capabilities for AI agents.

memorytypescriptagent +1
  • · Multi-sector cognitive memory: episodic, semantic, procedural, emotional, reflective
  • · Temporal knowledge graph with valid_from/valid_to for time-aware recall
  • · Self-hosted local-first storage via SQLite or Postgres, no vendor lock-in

MaiBot

5.9k · Python
Active A+

An LLM-based intelligent agent as a digital lifeform that values warmth, authenticity and genuine connection, with long-term memory and personalized conversation.

memorypythonagent +1
  • · Pursues natural human conversational style, rejecting mechanical GPT-like replies with casual varied chat
  • · Imitates different speaking styles in group conversations, autonomously learns slang and in-group language
  • · Long-term memory based on psychology personality theory, continuously building understanding of user preferences

Memori

16.3k · Python
Active A

Agent-native memory infrastructure that turns agent execution and conversation into structured, persistent state with an LLM-agnostic memory layer, MCP integration, and Python/TypeScript dual SDK support.

memoryagentrag +3
  • · Agent-native memory infrastructure converting conversations and execution into structured persistent state
  • · LLM-agnostic design supporting OpenAI, Anthropic, and any model provider
  • · Python/TypeScript dual SDK with MCP one-command integration for Claude Code/Cursor/Codex