agentmemory: give your AI coding agent a brain that survives the session

Every time you open a new Claude Code or Cursor session, your agent starts over. It asks what language you're using, what the architecture looks like, which patterns you prefer. You paste the same boilerplate context into CLAUDE.md, run the same /reset, retype the same "we use Postgres, not MySQL" correction. If you run several agents in parallel across multiple workspaces, the problem scales linearly.

agentmemory is an open-source local MCP server that fixes this 1. It records what your agent does during each session, compresses the observations with an LLM, and injects the right context back when the next session starts — automatically, without you touching a configuration file.

It hit #1 on GitHub Trending this week with 6.9K stars, accumulating roughly 3,000 stars in its first three days 2.

Install

npx @agentmemory/agentmemory

That's the entire setup. The server starts on port 3111. Open http://localhost:3113 to see the memory dashboard. No global install required, no Postgres, no Redis — the only dependency is SQLite (bundled).

To wire it into Claude Code, run this once after the server is up:

/plugin marketplace add rohitg00/agentmemory
/plugin install agentmemory

For Cursor, Gemini CLI, OpenCode, and any other MCP-compatible agent, add this to your agent's MCP config:

{
  "mcpServers": {
    "agentmemory": {
      "command": "npx",
      "args": ["-y", "agentmemory-mcp"]
    }
  }
}

A real scenario

You're mid-sprint on a monorepo. Claude Code helped you debug a tricky async race condition in the payment service yesterday. Today you open a fresh session and ask it to review a new PR in the same service.

Without memory, the agent has no idea what you worked on yesterday, which file patterns you established, or that you already ruled out a certain mutex approach.

With agentmemory running, the session-start hook fires automatically. The server retrieves the compressed episodic memory from yesterday's work — "debugging async payment service, ruled out mutex approach in checkout.ts, settled on retry backoff" — and injects it as context before your first prompt lands. The agent picks up roughly where it left off.

The retrieval uses a three-way hybrid: BM25 keyword matching, vector cosine similarity, and knowledge graph traversal, fused with RRF ranking 1. That combination outperforms BM25-only (the method behind CLAUDE.md static files) on the LongMemEval-S benchmark: recall@5 is 95.2% vs. the next-best alternative's 68.5%.

How the memory pipeline works

agentmemory organizes observations into four tiers — working → episodic → semantic → procedural — each corresponding to a different time horizon. Working-layer entries (raw tool call observations from the current session) get compressed by an LLM into episodic summaries. Recurring patterns graduate into the semantic layer as facts. Decision patterns and workflows harden into the procedural layer.

Twelve hooks capture events automatically across supported agents, so nothing requires manual /remember calls. Memory entries follow an Ebbinghaus-style decay curve: frequently accessed items strengthen, stale ones fade.

The entire operation costs roughly $10 per year in LLM tokens with a remote embedding provider, or $0 with a local model (all-MiniLM-L6-v2 works out of the box).

Star momentum

Metric	Value
Current stars	~6.9K
Stars in first 3 days	~3,000
Stars gained this week	#1 fastest-growing repo on GitHub
First appeared on trending	~May 13, 2026
Language	TypeScript
License	Apache-2.0

The LinkedIn deep-dive from AlphaSignal noted "3,000 stars in ~3 days" at launch 3, and a Reddit thread listed it as a replacement for Mem0's $50/month subscription 4.

The honest caveat

The HackerNews discussion around agent memory tools is not uniformly positive. A thread titled "Coding agents don't really need memory" 5 argued that git history, rules files, and documentation already serve this purpose more transparently. The top critique: opaque automatic context injection can cause agents to act on stale or wrong assumptions without you realizing it.

This is a real tradeoff. agentmemory addresses it partly through the memory dashboard at localhost:3113 — you can inspect, edit, and delete specific memories before they influence a session. But if you want full, explicit control over what your agent knows, a well-maintained CLAUDE.md is still a defensible choice.

agentmemory fits best when you run long-lived projects where the agent genuinely accumulates useful decisions over weeks, and where manually keeping a rules file current is the thing you keep forgetting to do.

github.com

rohitg00/agentmemory

#1 Persistent memory for AI coding agents — Apache-2.0, TypeScript, no external dependencies

リンクプレビューを読み込んでいます…

agentmemory: give your AI coding agent a brain that survives the session

Install

A real scenario

How the memory pipeline works

Star momentum

The honest caveat

rohitg00/agentmemory

参考ソース

rohitg00/agentmemory