AgentRecall
Show HN: I made an open-source memory layer for agents
Last verified:
What is AgentRecall?
AgentRecall is an open-source memory SDK for AI agents that provides persistent, intelligent memory across sessions. It solves the critical problem of AI agents forgetting everything between conversations by offering graph-based memory relationships, semantic search, and AI-powered processing. The system uses Neo4j-powered relationship graphs to connect memories, Qwen2.5-7B for automatic entity extraction and relationship detection, and vector embeddings plus full-text search for context-aware retrieval.
Key features include Graph Memory for traversing relationships between memories, AI Memory Processing that automatically extracts entities and categorizes memories without manual tagging, Semantic Search that finds memories by meaning rather than keywords, Cloud API with RESTful endpoints and authentication, Multi-Agent Support with isolated memory namespaces for each agent, and Bring Your Own Key flexibility for local or cloud deployment. The SDK works with Claude Code for developers, OpenClaw for personal assistants, and Hermes Agent for CLI power users through a single MCP server.
AgentRecall is designed for developers building AI agents, teams working with specialized agents, coding agents that need to remember architectural decisions and debugging context, personal AI assistants that should know user preferences and routines, and CLI power users who want plug-and-play memory without Neo4j setup. It supports both self-hosted deployment with full control and cloud hosting with zero setup.
The tool is open-source under MIT license, self-hostable with no vendor lock-in, and works with Claude Desktop via MCP. Users can install via npm or Python pip, initialize with an API key, and start storing memories in just three lines of code. The system scales to millions of memories and is built for production workloads.
AgentRecall pricing
Pricing model: Freemium
AgentRecall offers two pricing tiers. Free tier: $0 forever with 1,000 memories, basic semantic search, single agent, community support, and local mode (BYOK). Pro tier: $9/month with unlimited memories, graph memory (Neo4j), AI memory processing, smart semantic search, multi-agent support, cloud API access, and priority support. Self-hosted mode is free with MIT license, no API key needed, unlimited memories and agents, works with Docker or bare metal. Cloud mode provides hosted API with free tier of 1,000 memories and Pro at $9/month for unlimited.
AgentRecall pros
- Open source with MIT license - use for anything
- Self-hostable on your own infrastructure
- No vendor lock-in
- Free tier with 1,000 memories
- Neo4j-powered graph memory for relationship tracking
- AI-powered automatic entity extraction and categorization
- Semantic search by meaning not keywords
- Multi-agent support with isolated namespaces
- Same SDK works for both local and cloud mode
- Works with Claude Code, OpenClaw, and Hermes Agent
- MCP server support for easy integration
- Docker or bare metal deployment options
- No API key needed for self-hosted mode
- Unlimited memories and agents in self-hosted mode
- Pro plan at $9/month for unlimited cloud usage
- RESTful API with authentication and usage tracking
- Scales to millions of memories for production
- Three lines of code to get started
- Automatic memory processing - no manual tagging needed
- Cross-agent query capability when needed
AgentRecall cons
- Cloud mode requires API key and internet connection
- Free tier limited to 1,000 memories and single agent
- Self-hosted requires Neo4j setup for local mode
- AI processing depends on Qwen2.5-7B model
- Cloud Pro plan costs $9/month after free tier
- Limited community support on free tier
- Neo4j dependency for graph memory features
- Requires model setup for local self-hosted deployment
- Cloud mode has usage tracking and limits on free tier
Frequently asked questions about AgentRecall
What is AgentRecall?
AgentRecall is an open-source memory SDK for AI agents that gives them persistent, intelligent memory across sessions. It provides graph relationships, semantic search, and AI-powered processing so every conversation builds on the last instead of agents forgetting everything between sessions.
How do I install AgentRecall?
Install via npm with 'npm install agentrecall' or via Python with 'pip install agentrecall'. Then initialize with your API key in cloud mode or configure local mode with your Neo4j instance and model. You can start storing memories in just three lines of code.
What is the difference between cloud mode and local mode?
Cloud mode uses AgentRecall's hosted API with an API key, getting started in 2 minutes with no server to manage. Local mode runs on your own infrastructure with your Neo4j database and models, requiring no API key, with unlimited memories and agents. Both use the same SDK and API - the only difference is where your data lives.
Is AgentRecall open source?
Yes, AgentRecall is open source under MIT license. You can use it for anything, self-host it on your own server with full control, no costs, and no limits. Your data never leaves your infrastructure in self-hosted mode.
WhatAI model does AgentRecall use for memory processing?
AgentRecall uses Qwen2.5-7B to extract entities, detect relationships, and auto-categorize every memory. This AI processing happens automatically when you store memories, enriching your agent's knowledge base without requiring manual tagging.
How does semantic search work in AgentRecall?
AgentRecall uses vector embeddings plus full-text search to find memories by meaning, not just keywords. This allows your agent to understand context and retrieve what's truly relevant. You can set limits and thresholds for search results.
Can multiple agents use AgentRecall?
Yes, AgentRecall supports multi-agent deployment where each agent gets isolated memory with its own namespace. You can also perform cross-agent queries when needed, making it perfect for teams of specialized agents working together.
What platforms does AgentRecall work with?
AgentRecall works with Claude Code for developers (persistent memory for coding sessions), OpenClaw for personal assistants (remembers user preferences, relationships, daily routines), and Hermes Agent for power users (plug-and-play CLI memory). One MCP server supports all three platforms.
How does graph memory work?
Graph memory uses Neo4j-powered relationship graphs between memories. Memories aren't just stored - they're connected. You can traverse memory relationships with depth and relationship filters, finding hidden connections across your agent's entire history.
What is included in the Pro plan?
The Pro plan at $9/month includes unlimited memories, graph memory with Neo4j, AI memory processing, smart semantic search, multi-agent support, cloud API access, and priority support. This is the most popular plan for users who need more than the free tier's 1,000 memories and single agent.