rohitg00/agentmemory

⭐ 3,810  ·  TypeScript  ·  GitHub Repo

#1 Persistent memory for AI coding agents based on real-world benchmarks

agentmemory agents ai claude claudecode codex copilot cursor

1-Sentence Summary

Persistent, searchable memory for 15+ AI coding agents with 95% retrieval accuracy and zero external databases.

🔥 Key Capabilities & USP

• Universal Agent Memory: Eliminates the pain of re-explaining architecture, bugs, or preferences across 15+ different AI coding agents (Claude Code, Cursor, Gemini CLI, Codex CLI, OpenCode, Cline, Goose, Aider, and any MCP/REST client). All agents share the same memory server, creating a unified context that persists across sessions and tools.

• 95.2% Retrieval Accuracy: Achieves 95.2% recall@5 on the LongMemEval-S benchmark using hybrid search that combines BM25 fallback with semantic embeddings. This means your agent finds the right memory when it needs it, not just similar-sounding text.

• Zero External Database Dependencies: Runs entirely in-process with local embeddings support for zero-cost operation. No need to spin up PostgreSQL, Redis, or vector databases—just npx @agentmemory/agentmemory and you're running.

• 51 MCP Tools & 12 Auto Hooks: Comprehensive tooling for memory management with automatic hooks that silently capture agent activity. The hooks system lets you instrument any agent workflow without manual memory calls.

• Token & Cost Efficiency: Uses 92% fewer tokens than full context pasting (~170K tokens/year vs 19.5M+) and costs ~$10/year compared to $500+ for alternative solutions. This is the #1 persistent memory solution based on real-world benchmarks, not just marketing claims.

Technical Architecture

Component	Description
Core Engine	Built on the iii engine with hybrid search (BM25 + semantic embeddings)
Embeddings	Supports local embeddings for cost-free operation; remote embeddings optional
API Surface	104 REST API endpoints + MCP server for agent integration
Memory Store	In-process storage with confidence scoring, lifecycle management, and knowledge graphs
Hooks System	12 automatic hooks that silently capture agent activity without manual instrumentation
Filesystem Connector	@agentmemory/fs-watcher package for file-change-driven memory updates
Viewer & Debugging	Built-in real-time web viewer for inspecting memory state + iii console for debugging

Quick Start Guide

# Install and start the memory server
npx @agentmemory/agentmemory

That's it. The server starts with zero configuration, providing an MCP server, REST API, and hooks system. Your AI coding agents can immediately connect and begin sharing persistent memory.

Pros, Cons & Use Cases

Pros

• Works across 15+ agents—unified memory for your entire AI toolchain
• 95.2% recall@5 retrieval accuracy on real-world benchmarks
• Zero external databases—no infrastructure overhead
• Token-efficient—92% fewer tokens than context pasting
• Cost-effective—~$10/year vs $500+ for alternatives
• Open source with 51 MCP tools and 12 auto hooks

Cons

• Requires initial setup—the memory server must be running for agents to connect
• Memory quality depends on agent activity capture—poor instrumentation yields poor recall
• May have overhead for very simple, single-session workflows where persistence isn't needed

Who should NOT use this?

• Single-agent, single-session users who never switch tools or revisit projects
• Developers allergic to any server setup—even a lightweight in-process server
• Projects with trivial context where re-explaining takes seconds, not minutes

Ideal Use Cases

• Multi-agent workflows where Claude Code, Cursor, and Gemini CLI collaborate on the same codebase
• Long-running projects where architecture decisions and bug fixes span weeks or months
• Team environments where multiple developers use different AI assistants but need shared project memory
• CI/CD pipelines where agents need to remember build configurations and test failures across runs

Community & Activity

With 3,810 stars and active development (last updated May 2026), agentmemory has clearly struck a nerve in the AI coding community. The project is rapidly gaining traction as developers realize that context-switching between AI agents is the new productivity bottleneck. The comprehensive feature set—51 MCP tools, 12 auto hooks, 104 REST endpoints—shows a team that's shipping fast and listening to the community. If you're tired of repeating yourself to every AI agent you use, this is the project to watch (and star).