AgentMemory: Persistent Cross-Session Memory for Claude Code and 16 Other AI Agents

Why This Is Trending Today

rohitg00/agentmemory picked up 655 stars on May 11 alone, landing it in the GitHub daily trending chart with 4.2k stars total. That kind of single-day velocity on a developer tooling repo is worth paying attention to - it usually means the repo solved a problem that a lot of builders had already accepted as unsolvable.

The problem here is agent amnesia. Every Claude Code session, every Cursor chat, every Codex run starts from zero. You re-explain the project structure. You re-describe the architectural decisions you already made. You re-paste the relevant context. Multiply that across hundreds of sessions on a single project and you are burning tokens, time, and cognitive load on information that should have been remembered automatically.

agentmemory is a direct answer to that. It is a TypeScript package that spins up a local memory server, hooks into your agent via 12 automatic Claude Code hooks or a 51-tool MCP server, and handles capture, compression, deduplication, and retrieval without you touching anything manually.

What It Does

agentmemory runs as a sidecar process alongside your agent. It captures session observations - decisions made, code patterns used, project context established - and consolidates them into a 4-tier memory architecture:

1. Working memory - raw observations from the current session
2. Episodic memory - summarized session logs
3. Semantic memory - extracted facts and patterns across sessions
4. Procedural memory - repeatable workflows and decisions

Retrieval uses a triple-stream search combining BM25 keyword matching, cosine-similarity vector search, and knowledge graph traversal. The three signals are fused via Reciprocal Rank Fusion to produce ranked results. On the LongMemEval-S benchmark, it achieves 95.2% retrieval accuracy at R@5.

The architecture is designed to be lightweight. It runs on SQLite - no external databases required. The embedding provider is configurable: local (via all-MiniLM-L6-v2, which requires no API key), OpenAI, Gemini, Voyage AI, Cohere, or OpenRouter. A real-time viewer at localhost:3113 shows live observation streams, a session explorer, a memory browser, a knowledge graph visualization, and session replay with timeline scrubbing.

Built on the iii engine - a runtime that replaces Express, Postgres, Redis, pm2, and Prometheus with a unified local stack - the whole thing runs as a single npx invocation.

Install and Try It

Zero config to start:

npx @agentmemory/agentmemory

That starts the memory server on port 3111 (REST API, 107 endpoints) and the viewer on port 3113. To run with sample data first:

npx @agentmemory/agentmemory demo

To build from source:

git clone https://github.com/rohitg00/agentmemory.git && cd agentmemory
npm install && npm run build && npm start

To connect to Claude Code, run the server in a separate terminal, then inside your Claude Code session:

/plugin marketplace add rohitg00/agentmemory
/plugin install agentmemory

Once installed, the 12 hooks fire automatically on session start, tool calls, file edits, and session end - capturing context without any manual remember commands. You can also import your existing Claude Code JSONL session history:

npx @agentmemory/agentmemory import-jsonl

The MCP server exposes 51 tools: memory_recall, memory_save, memory_smart_search, memory_compress_file, memory_sessions, memory_timeline, memory_graph_query, memory_team_share, and dozens more. Four skills ship out of the box: /recall, /remember, /session-history, and /forget.

Who Should Use It

Claude Code users working on long-running projects get the clearest benefit. If you have a project you return to weekly, or a codebase where you repeatedly re-explain conventions and architectural choices, agentmemory removes that overhead. The 12 auto-hooks mean you get memory capture from day one without changing how you work.

Multi-agent teams will find the memory_team_share tool and federation features relevant. You can share project memory across agent installations - useful when multiple agents work on the same codebase, or when you hand off context from one agent type to another.

Token-conscious builders - the repo claims 92% token savings compared to context-pasting approaches (approximately 170K tokens per year versus 650K+). That number depends heavily on how much context you normally paste, but the directional claim is plausible: retrieval pulls only what is relevant rather than injecting everything.

Developers using multiple agent platforms - agentmemory supports 16+ agents including Claude Code, Cursor, Gemini CLI, Codex CLI, Cline, Goose, Aider, Windsurf, and Claude Desktop. A single memory server can feed context to whichever agent you are using at a given moment.

Relation to the DevDigest Ecosystem

This repo connects directly to two things we track closely.

MCP tooling - 51 MCP tools is a serious surface area. The MCP server directory at mcp.developersdigest.tech tracks servers by category; agentmemory adds a dense memory-and-retrieval surface to any MCP-compatible agent, which is a different category from the filesystem, browser, or code-execution servers most people run first.

Claude Code skills and hooks - the skills directory at skills.developersdigest.tech documents the skills and hooks pattern for Claude Code. agentmemory is a production example of what that pattern looks like at scale: 12 hooks that fire across the session lifecycle, 4 skills that surface retrieval into the agent interface, and a plugin install flow that hooks into the Claude Code plugin marketplace. If you have been thinking about how to build durable memory into your Claude Code workflows, agentmemory is worth pulling apart as a reference implementation - not just as a tool to run, but as an example of how hooks and skills compose into a coherent system.

The broader agent-memory problem is also covered in AI Agent Memory Patterns, which goes deeper on the architectural tradeoffs between file-based, RAG-based, and context-compression approaches.

Honest Assessment

What works well: The zero-config path is genuinely low friction. One npx command and the server is running. The decision to use SQLite as the backing store - rather than requiring a local Postgres or Redis - is correct for a developer tool that should stay out of the way. The 827 passing tests and 118-file TypeScript codebase suggest this is not an MVP prototype. The support for 16+ agent platforms means the investment in setup is not locked to a single tool.

What to watch: The benchmarks (95.2% retrieval accuracy, 92% token savings) are self-reported and measured on LongMemEval-S, which is a controlled benchmark rather than real-world session diversity. Real-world retrieval quality depends heavily on how well your agent's observations get captured - sessions with very short outputs or highly technical specifics may not compress as cleanly as the benchmark suggests.

The plugin marketplace install path for Claude Code (/plugin marketplace add rohitg00/agentmemory) is newer surface area in the Claude Code ecosystem - if you run into friction there, the MCP server path (claude mcp add) is a more established alternative.

The 51 MCP tools is also a large surface to understand. Starting with memory_recall, memory_save, and memory_smart_search and ignoring the rest until you need them is the practical approach.

At 4.2k stars and Apache-2.0 licensed, this is early but real. It is worth watching over the next few weeks to see whether the retention metrics from actual users match the benchmark claims.

References

• rohitg00/agentmemory on GitHub
• iii engine (underlying runtime)
• LongMemEval benchmark
• MCP server directory - mcp.developersdigest.tech
• Claude Code skills directory - skills.developersdigest.tech
• AI Agent Memory Patterns - Developers Digest
• Best MCP Servers 2026 - Developers Digest