Why AI without memory starts from scratch every day

Anyone who has worked with an AI assistant for more than a day knows the feeling: yesterday you explained how your company writes quotes, and today it asks again. The model hasn't forgotten — it never remembered in the first place. Language models remember nothing between two conversations. Everything they know about you has to be handed to them again every time.

For a single question that is fine. For an assistant that works on the same project for weeks, answers customers or runs processes, it is the difference between a colleague and an intern who arrives for their first day every morning.

Context is not memory

You can put a lot of text into each conversation — the so-called context. But context is a desk, not an archive:

• it is limited — not everything fits, and the fuller it is, the slower and costlier the answers,
• it disappears when the conversation ends,
• it doesn't tell important from trivial — last month's decision and a passing remark weigh the same.

Real memory means the system itself stores what is worth remembering and finds it again when it is needed — without you pasting the same instructions every time.

Two kinds of data bank

In practice an AI needs two different things: knowledge that people write and maintain, and memory the agent builds itself from its work. We use two tools for that.

Obsidian — knowledge people write

Obsidian is a note-taking app where everything is a plain Markdown file in a folder on your computer (a vault). Notes link to each other, forming a web of knowledge that is easy to search — for people and for AI.

It holds what has to be correct and approved:

• decisions and the reasons behind them ("why we use no analytics cookies"),
• rules and conventions ("how we write URLs", "what every new page looks like"),
• instructions for recurring work — for example, how a new version of the site is released.

Because they are plain files, an AI agent reads them directly, and every change has a history just like code. This website, the blog you are reading included, has such a vault with a decision log: before the agent changes anything, it reads what has already been agreed.

Hindsight — memory the agent builds itself

Hindsight is an open-source long-term memory system for AI agents. It doesn't store conversations word for word; it extracts facts, people, concepts and time from them and links them into a graph. It works through three core operations:

• retain — store new information,
• recall — find relevant memories (searching in parallel by meaning, keywords, graph links and time),
• reflect — reason over everything remembered and answer a question that needs thinking, not just lookup.

Memory lives in banks: each is fully isolated — one per user, agent or project — so what the agent learns on one project never spills into another.

Hindsight holds what happens during the work: what was tried and didn't work, which bugs are already solved, what the client said in a meeting three weeks ago.

Why both

Obsidian is what has to be true — which is why people write and approve it. Hindsight is what happened — which is why the agent fills it itself, without anyone taking minutes. When the two disagree, good practice is to trust the documentation and check the current state, not to follow the memory blindly.

The result: an assistant that knows your rules, remembers what happened and doesn't ask the same question twice.

What to watch out for

• Passwords and keys never go into memory. Treat memory like any other database — and what is never stored can never leak.
• Personal data is processing under GDPR. If the agent remembers customers' names and emails, say so in your privacy policy and have a data processing agreement with the provider.
• One bank per client. Separate banks aren't just tidiness — they guarantee one client's data never ends up in an answer for another.
• Memory ages. What was true six months ago may not be any more. That's why important decisions live in the documentation, where someone maintains them.

AI without memory is clever but forever new. With good memory it becomes a colleague who knows how you work.