Memory: Why Notes and Git Became More Important Than Good Prompts

An LLM can be impressive and still not know tomorrow what mattered yesterday. That is not an insult. It is a working condition.

In the beginning, I left too much inside chats. Decisions, ideas, failed attempts, small lessons. Everything was scattered across conversations. That feels fine as long as you stay in the same thread. Later it becomes painful. You search. You half remember. You explain the same thing again.

The next step was not a bigger model. It was memory.

By memory, I do not mean mystical AI memory. I mean very ordinary things: notes, project overviews, decision records, Git repositories, clean handoffs. A place where the current truth lives. A place where a new AI session does not have to start from zero.

Obsidian became useful for me because it makes it easy to keep thoughts in Markdown. A project gets a page. A decision gets a date. A system gets a short description. Not perfect. Just findable. Git became important because it makes changes traceable. What changed? When? Why? What is the current state?

This changes AI work a lot. Instead of telling the model a long story every time, I can say: read the project overview, respect the rules, summarize what matters for this task, then work.

Good memory is not a lot of text. Good memory is the right text in the right place. One page with five clear rules is more useful than twenty pages of diary when I need to make a change. A short project index is more useful than a folder full of unsorted notes.

This theme later became concrete in public experiments such as claude-cortex and claude-sync. The idea behind them is simple enough for non-technical readers: an assistant becomes more useful when it does not have to rediscover your working context every time. Memory is not about making the AI "alive." It is about reducing repetition.

There is a trap, though. More memory is not automatically better memory. If you save everything, you create noise. If you save private data in the wrong place, you create risk. If you never update old notes, the assistant may confidently follow stale instructions. Memory needs hygiene.

For beginners, the practical version is simple: create one project page. Put the purpose, current status, important terms, open questions, and next steps on it. When you work with AI, give it that page as context. After the work, add what changed. That is enough at the start.

In normal words

Markdown is a simple text format for notes. You can write headings, lists, and links without needing a complex editor.

Git is a system that records changes over time. It lets you see what changed and go back if needed.

Handoff means a short summary that helps the next session, person, or assistant continue without starting over.

The effect is surprising. AI does not automatically become smarter, but it becomes less disoriented. So do you.