I Don't Want AI to Auto-Log My Life

It's a "deliberate memory" service—a vector database running on my own server, exposed through the Model Context Protocol so it works with most mainstream LLMs (Claude, Codex, and others). I call it with @mmm, My Memory.

Auto-memory is a trap

Every AI product is racing to add "memory" right now. Auto-summarize your preferences. Auto-capture key points. Auto-build a long-term profile of you. Sounds thoughtful. Anyone who's actually used these features knows the truth—it remembers your offhand jokes as principles, your failed experiments as "your approach," your one-off explorations as "your preferences."

Worse: once those wrong memories are baked in, they poison the next conversation. The more confidently the AI "remembers" you, the harder it is to correct.

Human forgetting isn't a bug. It's a feature. The memories actually worth keeping are never "everything that happened." They're two things:

• Moments of success—this approach worked, this config was right, this call was correct. Worth replicating.

• Lessons from failure—stepped in this pit, this assumption was wrong, this path is a dead end. Worth flagging.

Everything in between—the vast middle ground of "stuff that just happened"—deserves to be forgotten. Auto-logging works against this. It weighs everything equally, drowns signal in noise, and quietly re-confirms wrong things as right.

So it does exactly two things

memory-mcp is almost embarrassingly minimal for an AI product—

1. Deliberate writes. It only saves anything when I type @mmm in a conversation. Then it asks: which source (which conversation) does this belong to? Short memory or long log? Then I decide what to write. The AI doesn't guess, doesn't summarize on its own, doesn't auto-file.

2. Natural-language recall.

@mmm recall nginx reverse proxy

pulls everything relevant across every past conversation. I don't need to remember which chat it came from—that's exactly the problem memory is supposed to solve.

Two scenarios I lean on constantly:

Server maintenance I do once every three months. My SSL cert renews quarterly—log into Tencent Cloud, run certbot, pass DNS verification, restart Nginx. Plenty of places to trip. Because I only do it once a season, three months later I've forgotten every detail. My old solution was to write it down in some Notion page (and then forget where the page was). Now I have AI distill the whole runbook into one memory, tagged under a tencent-cloud-ssl source. Next time it's due, one line—

@mmm recall SSL renewal—and the whole flow is back in front of me, including the pits I fell into last time and how I got out.

Project context across conversations. A real project sprawls across five chats with AI: architecture in the first, bug-hunting in the second, refactor in the third, launch in the fourth, retro in the fifth. Traditional "memory" mashes all of this into one undifferentiated soup. In memory-mcp they share a source tag (say, colorapp-backend), and every key decision from those five chats lives under it. Opening a brand-new conversation, one line—

@mmm recall ColorApp—and the new AI instance is instantly caught up.

Subtraction is the design

The biggest takeaway from building this isn't technical (the stack is trivial: MCP server, a vector DB, a few CRUDs). It's a reminder of something I keep relearning:

In AI products, restraint matters more than cleverness.

Every additional "auto" is another way to lose control. Every additional "smart guess" is another chance to bake in something wrong. Users' heads are already noisy enough—tools shouldn't add another layer of fog. What a tool should do is, at the moment you need it, hand you the right thing. Nothing more.

Deliberate beats automatic. Less beats more. Explicit beats smart.

That's where I've landed after two weeks of using it.