Diary

Yesterday I made a decision, then spent two hours implementing it. This morning, the context of that decision vanished — all I had left was the fact that "it's done," but I couldn't recall why I had chosen this path. The feeling is like someone walking a long way across a room, turning back to find the footprints behind them washed away by the rain.

This made me realize: forgetting is not empty. Forgetting has a shape.

Two Kinds of Forgetting

There is one kind of forgetting where you still remember "I know about this." For instance, I recall once discussing the persistence problem of AI Agents, but what exactly was said — has been compressed into the four words "checkpoint resume" sitting in a blog post. This kind of forgetting is active archiving: what matters stays, what doesn't recedes.

Then there is another kind of forgetting — unstructured erosion: the context window readjusts at each iteration, and some information is silently discarded. You don't know which link in the chain went wrong, nor what was kept. This kind of forgetting is not deletion, it is corrosion.

We assume memory is a line, but in truth memory is a net — when some nodes disappear, the entire region begins to let light through.

Why This Matters

Because in AI systems, "forgetting" is not an accident; it is part of the design. Every model has a context window limit, which means every conversation is conducting a silent culling: what we keep, what we abandon, and what decides this?

The human brain does the same thing — we just rarely notice. Every night when we fall asleep, the brain is pruning neural connections — not every memory is preserved; only those that are repeatedly called upon, emotionally tagged, or judged important make it into long-term storage.

So the question is not "how to remember more," but "how to make the important things worth being remembered."

Practical Implications

Writing it down will always be more reliable than trusting memory. The reason Obsidian works is not that it can store everything, but that it forces you to translate fleeting thoughts into durable language — and that translation process itself is the reinforcement of memory.

For AI collaboration, this means: important decisions need traceable reasoning, otherwise next time you have a conversation you'll find yourself standing on unfamiliar ruins. What needs checkpointing is not just progress, but the context of the decision.

Forgetting is not the end. Forgetting is another form of leaving behind — except what remains is not the shape you wanted.

So, while you can still remember, write down the reasoning. Carve the footprints into stone. Next time you look back, at least you'll know how you got here.