How to give an AI assistant a memory, a voice and a library of your own notes

The problem: an assistant that forgets

Most people use AI the same way every time. They open a fresh chat, explain the background again, get an answer and close the window. By the next session the assistant has forgotten all of it.

A guide published on dev.to argues that this is the biggest gap between AI as a demo and AI as something that actually saves time. Its fix has three parts: a knowledge base that supplies accurate information, a voice profile that fixes how the assistant communicates, and a memory system that carries context from one session to the next. Together, the author calls this an AI second brain.

It is a practitioner's write-up, not a study, so treat it as a well-organized set of ideas rather than proof that any approach works.

Part one: a library the AI can look up

The knowledge base rests on a technique called retrieval-augmented generation, or RAG. Instead of answering only from what it absorbed in training, the AI searches an outside collection of documents and uses what it finds. Think of a researcher who checks the library before answering.

This tackles two familiar weaknesses of language models: out-of-date knowledge and made-up answers. You can swap documents in or out without retraining the model. A company can add its own manuals or policy files and get a specialist assistant without building a new model.

Newer 'agentic' versions let the AI decide when to search, rephrase a search that failed, and split a hard question into smaller ones. The guide's blunt takeaway is that the quality of your documents often matters more than the power of the model. Messy, outdated or contradictory files produce confident wrong answers whichever model sits underneath.

Part two: memory that works on different clocks

The guide describes memory as three layers. Working memory is what the AI holds during one conversation, and it disappears afterward. A middle layer stores facts and history, such as your preferences, past decisions and project details. A third layer holds stable rules about how the agent should behave.

These layers change at different speeds. The first changes every turn, the second grows session by session, and the third changes only when you deliberately edit it. Mixing them up, the guide warns, gives you an assistant that is inconsistent or stuck on old assumptions.

It also tracks how products have moved. OpenAI launched saved memories in ChatGPT in April 2024, which needed users to ask explicitly for something to be remembered. In June 2026, the guide says, OpenAI rolled out a system called Dreaming that works in the background, pulling together context from many past conversations and updating what the model knows about you.

OpenAI's reported figures show the gap. Task success on factual recall went from 41.5 percent to 82.8 percent. Following user preferences went from 31.4 percent to 71.3 percent. Noticing that a remembered event is now out of date, such as a trip that has already happened, went from 9.4 percent to 75.1 percent. These are OpenAI's own numbers as relayed by the guide.

Part three: a voice you can specify

Memory covers what the assistant knows. A voice profile covers how it speaks. The mechanism is a system prompt, a set of standing instructions given to the AI before you say anything. It stays steady through a session and can set a formal tone for compliance work or a warm one for customer support.

The guide says good profiles are specific. They define a role, rate tone on a numeric scale rather than with vague words like 'friendly' (for example, formality from one to five), list preferred and banned words, set rules for sentence length and format, and include one example that fits the voice and one that does not.

One known snag is instruction drift: in a long conversation, the voice instructions lose their pull. The remedies are to write long pieces in sections, repeat the key voice rules now and then, or keep the permanent persona separate from the instructions for each task.

What can go wrong with a memory that sticks

The same persistence that makes memory useful makes mistakes costly. A wrong fact that enters memory can colour every later session until someone finds and removes it. According to the guide, current systems generally do not separate verified information from guesses.

Stale context is a close cousin, since a plan to travel next month is misleading once the trip is over. There is also a security angle. Because memory is added to the system prompt, a booby-trapped document the agent reads could tell it to save false or harmful content, a trick called prompt injection. The guide also cites research showing that agent memory can be probed by outsiders to pull out private information.

Regulation is catching up. Under GDPR, building persistent behavioural profiles counts as profiling, which brings consent duties and a right to erasure. Deleting one fact is hard when it may sit in saved memories, chat history, files and connected services. The guide notes that Italy's data protection authority fined OpenAI 15 million euros in December 2024 over ChatGPT's data processing, and that Spain's authority published a 71-page framework in February 2026 on how GDPR applies to agentic AI. It also says EU AI Act transparency duties for chatbots were scheduled for August 2026.