---
description: Providing task rules and project knowledge selectively instead of storing everything indiscriminately.
url: https://www.schmalzl.tirol/en/blog/2025-10-18-ki-agent-gedaechtnis-skills
lang: en
date: 2025-10-18
deutsch: /blog/2025-10-18-ki-agent-gedaechtnis-skills.md
---

# What an AI agent should remember

_18 October 2025 · Archive date_

When a task comes up more often, you do not want to explain the same basics every time. An AI could, for example, remember that a small sample project uses German labels and that new features should first be shown with invented data.

These are useful working rules. It does not follow, however, that every message belongs in a permanent memory. The selection is part of designing the system.

## Three different kinds of knowledge

The collection needs three areas: rules, tasks and sources. A rule might be: "Always give measurements with a unit." The current task concerns a particular model. The source describes a folding technique and stays linked to its origin.

A completed task does not need to keep turning up as a current instruction. Equally, a guess should not quietly become a permanent rule. A note such as "unchecked" or "replaced by a newer version" keeps the state understandable.

Skills can be seen as working instructions that are provided selectively. What matters is the right scope: a guide for images needs different information from one for tables. Loading everything at once does not necessarily make a task clearer.

## Memories must stay correctable

A usable memory shows where a piece of information comes from, when it was checked and how it can be changed or removed. Otherwise the AI works with assumptions that nobody can trace any more.

The simplest way in is therefore a short, readable project file with a few rules. Only when a specific task needs more knowledge are further documents added. That way it stays clear what the system knows and which assumptions it uses for its next answer.

## Original finds

These links were the starting point for this post. The commentary in the text is mine; the linked posts are not my own tests.

- [Skills: loading knowledge to match the task](<https://www.linkedin.com/posts/danielmiessler_business-ai-is-the-automation-of-intelligence-activity-7385100506261676032-otl3>)
- [What an AI agent should remember](<https://github.com/rohitg00/agentmemory>)
- [The architecture around a language model](<https://www.linkedin.com/posts/a-harnessed-llm-agent-and-why-llm-is-a-share-7465296752481427456-A8LI/>)
- [Claris Docs Skills for targeted help articles](<https://www.linkedin.com/posts/ask-an-ai-coding-agent-a-filemaker-question-share-7475181752890667008-nhlu/>)
- [Building an agent knowledge base from permanent profile knowledge and continually added processes and content.](<https://www.linkedin.com/posts/felixschlenther_so-habe-ich-die-wissensbasis-f%C3%BCr-unser-ki-agenten-team-activity-7387011724320858112-CJUG>)
- [Setting up Cowork with profile, style, project, template and output folders.](<https://www.linkedin.com/posts/hruthikvarma_how-to-set-up-claude-cowork-in-1-hour-the-share-7448085758554398720-ba9_>)
- [Cowork setup with profile, style, company and template files; your own company information stays private.](<https://www.linkedin.com/posts/claude-claudecowork-aiproductivity-share-7451875368774369280-qhcw>)
- [Advert for a permanent ChatGPT setup with personal context and reusable rules.](<https://www.linkedin.com/posts/canerman-ai_h%C3%B6r-auf-chatgpt-jeden-tag-neu-zu-erkl%C3%A4ren-share-7454455207167578112-sRaF>)
- [noirdoc replaces personal data in documents with placeholders before LLM processing.](<https://www.linkedin.com/posts/antoniomaiolo_noirdocde-jetzt-als-open-source-cli-ugcPost-7454816612823662592-NY_T>)

> **About the date:** This post was written on 11 October 2026. The archive date corresponds to the earliest related find in the collection, not to a publication at that time. Technical context and sources added later refer to the time of writing.