Agentic Coding Summit #1: Why Does My AI Forget Everything?

Welcome & Introduction

Opening remarks for the first Agentic Coding Summit.

GitHub Copilot Memory: From Black Box to Memory Bank

GitHub Copilot Memory promises to give your AI a long-term memory. But what does it actually remember, and who decides? In this talk, I explore the official Copilot Memory feature, a cloud-hosted black box that learns automatically but offers no user control, and contrast it with the Memory Bank pattern: a set of Git-tracked Markdown files that Copilot reads and updates with every task. No plugins, no external tools. Just a copilot-instructions.md and a folder of Markdown files that give your AI the context it needs to stop forgetting.

Dr. Matthias Liebeck

Meet Jarvis, My Vibe Manager

The AI workflow that keeps me sane, organized, and productive as a founder. The exact setup with Memory + Skills + Claude Code + Linear. I never touch my project management system and just talk to it. My agent skills ensure the process is remembered and get better with every iteration. It turns out this system is also a perfect task board for my OpenClaw instance "ChristAIna" to take work of my shoulders.

Ben Sufiani

Break

I think my AI has ADHD

The task was completed even though the goal hasn’t been reached yet? Suddenly, things were implemented that weren’t planned? Our agreement was forgotten? And why is there Python in my JavaScript?

Since the AI can’t go to therapy, it needs to be helped in other ways. With memory optimisation and self-monitoring routines.

This talk demonstrates how to set up a governance and control system using OpenAI Codex.

Tim Dorbandt

Would You Still Be You Without Your Memories? Onboarding Chlawe, My OpenClaw Agent

Every morning, most AI agents wake up as strangers. Their skills are intact, but they have no idea who they are working with or what happened yesterday. That is not a tool problem, it is an identity problem. I built a 4-layer memory system for my OpenClaw agent, Chlawe, treating her like a new team member that deserves proper onboarding. The system combines always-injected working memory, a scored knowledge graph where facts gain importance through use and fade over time, a self-cleaning archive, and a nightly self-improvement loop where Chlawe reviews her own mistakes and updates her instructions. Facts follow a maturity lifecycle from draft to validated to core and a compound scoring formula ensures the most trusted knowledge surfaces first. Everything is plain Markdown and fully git-friendly. I will demo it live and share a practical implementation guide you can apply to your own agents.

Anna Lübken

Break

Lies, Confusion & Amnesia: Giving AI a Personal Memory

AI assistants are incredibly powerful — but they also forget, drift, simplify critical details, and sometimes confidently lead humans into shared confusion.

In this lightning talk, I’ll share a personal journey from early “vibe coded” AI projects to experimenting with structured memory systems, context engineering and multi-agent workflows.

Why do AI systems often break down after initial success?
Why do they lose goals, architectural consistency and context over time?

And why is the real challenge often not intelligence — but memory?

Using practical examples from experiments at ORAYLIS and private side projects, I’ll show how lightweight approaches like structured Markdown memory, reusable AI skills, development guidelines and personal MCP servers can dramatically improve reliability and consistency.

But the talk goes one step further:

Once AI systems start remembering, optimizing and operating increasingly autonomously, the discussion is no longer only about better prompts or larger models — it becomes a question of context control, alignment and ultimately human sovereignty in the AI age.

The result is not perfect AI.
But AI systems that become more understandable, controllable and trustworthy for complex development and agentic workflows.

Jens Kröhnert

Knowledge That Survives the Reset: A File-Based Approach to AI Context

AI coding tools are getting better at remembering context across sessions. But the session reset problem points at something worth solving regardless: how do you treat working knowledge as a durable artifact rather than something that lives only in conversation?

Skill files and custom slash commands in Claude Code are two approaches to this question -- structured, version-controllable files that encode methodology, style, and workflow in a form the model can read on demand. They work alongside native memory features, but they also survive without them.

This talk shares a practitioner's experience building this kind of architecture, including what the failure mode looks like when you skip it, and what changes when you start treating your working knowledge as a corpus to maintain rather than a conversation to repeat.

Daina Bouquin

Roundtable: AI Memory: Solved Problem or Still Broken?

All speakers discuss trade-offs, gaps and what they want to see next.

End