First, What This Actually Is
It's a small free program that runs in the background on your computer. It sits between you and Claude. Every time you and Claude work on something, it quietly takes notes — what your project is, what decisions you made, what your style looks like, what files matter.
Next time you start a fresh session with Claude, AgentMemory hands it the right notes before you say a word. Claude already knows where it left off. You don't have to re-explain anything.
That's the whole product. A memory layer Claude can read from automatically.
Claude Has Goldfish Memory By Default.
Anyone who uses Claude for real work has felt this. You spend 20 minutes setting up context. "Here's my stack. Here's how I name files. Here's the decision we made yesterday about auth." You finally get going. You ship the thing.
Next morning you open a new session. Claude has forgotten everything. Twenty minutes of context-setting again. Same conversation, same explanations, same wasted tokens. Times every project. Times every session. Times forever.
AgentMemory ends that loop.
92% Drop In Tokens. Here's What That Means In Dollars.
The big number you'll see everywhere: 92% reduction in context tokens. Translated into something you can feel: roughly $10/year in token costs instead of $100+ if you tried to dump everything into the conversation manually.
But the bigger win for most people isn't cost. It's the $20 Claude plan. If you keep hitting your usage limits halfway through the day, this fixes it. Less context dumped into every message = more messages you get to send before running out.
One honest caveat:the 92% number compares AgentMemory against naive "dump everything into a big text file" approaches. If you weren't doing that to begin with, you won't see exactly 92%. You'll still see a massive drop — but treat the headline number as a ceiling, not a guarantee.
Karpathy Sketched It. AgentMemory Built It.
Andrej Karpathy (ex-OpenAI co-founder, ex-Tesla AI lead) posted about something he called the "LLM Wiki Pattern." The idea was simple: instead of dumping all your project info into the prompt every time, keep a structured wiki of facts the model can look up only when needed.
AgentMemory is the production-grade version of that idea. It adds smart memory tiers (short-term working memory, long-term knowledge, behavioral patterns), automatic capture, hybrid search, and a clean dashboard. You don't maintain anything by hand. The tool does the wiki for you.
Open A Terminal. Paste This. Done.
Open Terminal (Mac) or PowerShell (Windows). Paste this:
// run once
Hit enter. The first time, npm will ask if it's OK to install — say yes. AgentMemory boots up its server on your machine and starts listening for any AI coding agent that asks for memory.
If you prefer to install it permanently instead of running it fresh each time:
Don't have npm installed? Install Node.js first from nodejs.org (pick the LTS version). Then run the AgentMemory command above.
You Don't Have To. It's Automatic.
This is the part that surprises people. Once AgentMemory is running, Claude Code finds it automatically. No config files. No flags. No extensions to install in Claude. Just open Claude Code like normal and start working.
Behind the scenes, Claude Code's lifecycle hooks fire as you work — session start, every tool use, every message, session end. AgentMemory listens to those hooks, decides what's worth remembering, and writes it to a local SQLite database. Sensitive stuff (API keys, passwords) gets filtered out automatically.
The Magic Shows Up In Session Two.
With AgentMemory running, open Claude Code and do real work for 10 minutes. Build something. Have a conversation. Make a couple of decisions. Tell Claude things about your project.
Close the session completely. Open a new one. Ask:
Claude will pull up the relevant memories AgentMemory captured. Your stack, the decisions you made, the files you touched, the patterns you like. No re-explaining. The session you closed five minutes ago is now permanent context.
See What Claude Knows About You.
Open your browser and go to:
That's the AgentMemory dashboard. You'll see every memory the tool has saved, organized by type — facts about your project, your preferences, decisions you made, patterns it noticed.
From the dashboard you can search memories, edit them by hand if Claude got something wrong, pin important ones so they always inject, or delete anything you don't want stored. You own the data. It lives on your machine.
One Memory Store. All Your Tools Read It.
This is the underrated piece. AgentMemory doesn't just talk to Claude Code. It works with all of these too:
- Cursor, Windsurf, Cline, Roo Code — the popular AI editors.
- OpenAI Codex CLI, Gemini CLI — if you use other model providers.
- Hermes Agent, Goose, Aider, OpenCode — the autonomous agents.
Same memory, every tool. Build context in Claude in the morning, switch to Cursor in the afternoon, the memory follows you.
Bookmark These.
- Official site: agent-memory.dev
- GitHub repo: github.com/rohitg00/agentmemory
- Install Node.js (if you don't have it): nodejs.org
- Download Claude Code: claude.com/download
- Karpathy's LLM Wiki post (the inspiration): x.com/karpathy
One Command. Permanent Memory. Way Cheaper Claude.
That's the whole upgrade. Open a terminal, run one command, AgentMemory takes care of the rest. Your $20 Claude plan starts to feel like the $100 plan. Your projects build on themselves instead of starting from zero every morning.
Tag me when you've got it running. @tenfoldmarc.