AI Daily: Is the Long-Term Memory Problem for AI Agents Solved? New Visions from Adobe and Anthropic

Facing the tidal wave of new AI news every day, do you occasionally feel information overload? But seriously, today’s updates are worth stopping to take a good look at. From how developers are letting AI “remember” longer workflows, to how creatives are taking back control of AI, these technical advancements are quietly changing the way we interact with our tools.

We are no longer just giving simple commands to chatbots; we are building more complex, persistent collaborative systems.

This article will take you through how Anthropic is solving the “forgetfulness” problem for AI engineers, how Adobe intends to reshape the creative process with node-based editors, and the new breakthroughs from Perplexity and Google Gemini in personalization and learning.

Anthropic Proposes New Architecture: Letting AI Agents “Hand Over Shifts” Like Human Engineers

A headache often discussed in developer circles recently is that when we ask AI to handle a complex task taking hours or even days, it often “gets lost.” Because current AI models are limited by the Context Window, every new conversation is like a new employee who just started work and has no idea what happened yesterday. For those wanting to build Long-running Agents, this is simply a nightmare.

Anthropic’s engineering team clearly realized this too, and they have just released a highly valuable research piece titled Effective harnesses for long-running agents. This is not just a technical document; it’s more like an operations manual for developers.

Dual Agent Model: Initializer & Coding Agent

They drew inspiration from the “shift work” system of human software engineers. Since a single AI cannot handle everything at once, break it down. Anthropic proposes a “dual solution”:

1. Initializer Agent: This is like a project manager or architect. Its job is to set up the environment during the first run, write an init.sh script, and create a claude-progress.txt file to track progress. It is also responsible for the first Git commit, telling everyone: “Hey, this is our starting point.”
2. Coding Agent: This is the engineer who actually does the work. In each subsequent session, it is responsible for incremental development. Most importantly, it must leave “clear handover documentation” before finishing its work.

Curing AI’s Bad Habit of “Biting Off More Than It Can Chew”

Interestingly, Anthropic found that models of Claude’s caliber have two common failure modes. First, it tries too hard to show off, attempting to write the entire App at once (One-shot), which often results in getting cut off halfway due to context limits, leaving a mess. Second, it sometimes gets overconfident, glancing at things and thinking “Okay, I’m done,” when the functionality actually doesn’t work.

To combat these issues, they introduced several key mechanisms:

• Mandatory Feature List: Have the Initializer Agent write a detailed feature_list.json first, listing all features and marking them as “failed”. The Coding Agent can only focus on changing the status of one feature to “pass” at a time.
• Environment Cleanup and Testing: Require the AI to run tests after every code modification, just like a human engineer. If the test doesn’t pass, it can’t say it’s finished. This not only reduces bugs but also allows the next AI taking over to continue working in a clean environment.

This method ensures AI no longer writes code based on luck, but with discipline and method.

Adobe Project Graph: Refusing the Lottery, The “Node-based” Revolution of Creative Workflows

Having discussed hardcore programming, let’s look at the creative field. If you’ve used generative AI for drawing, you’ve surely felt this: writing a Prompt is like buying a lottery ticket; you never know what the next image will look like. For professional designers, this randomness is unacceptable.

Adobe has just released Project Graph, which might be exactly what creative workers have been waiting for. Simply put, it tries to stuff the powerful capabilities of AI into a “controllable” box.

From “Guessing” to “Designing”

Project Graph is a node-based visual editor. If you’ve used Blender’s material nodes or Unreal Engine’s blueprints, this interface will be familiar.

• Visualized Flow: You can string together Photoshop functions, AI models, and various effect tools like a connect-the-dots game. This means you can precisely control every step, rather than praying to a chat box.
• Tool Encapsulation and Sharing: This is the coolest part. Once you’ve designed a complex workflow (e.g., auto-remove background -> color grade -> add shadow -> generate background), you can “package” it into a simple tool. Your colleagues don’t need to understand the complex nodes behind it; they just need to click a button to use the workflow you designed.

This represents a core philosophy of Adobe: AI should not replace the creative process, but should become material in the hands of creators. This modular, reusable design is the kind of AI application that has the potential to enter professional production lines.

Perplexity’s Memory Upgrade: It Finally Knows Who You Are

The next update might seem minor, but for those who use AI search every day, it’s very thoughtful. Perplexity announced that their system can now “remember” your conversation threads and interests.

What does this mean? Previously when using AI search engines, every new window was like talking to a stranger. But now, Perplexity can call upon memory across models and search modes.

• Context Across Time: You can continue a conversation from weeks ago without re-explaining the background.
• Personalized Answers: If you previously told it you are a developer using Python, the next time you ask a programming question, it won’t give you a Java example.

This “long-term memory” capability is a key step for AI assistants evolving from “tools” to “partners.” It reduces the time we spend repeatedly inputting background information, making information acquisition smoother.

Google Gemini Making Learning “Alive”: Interactive Image Features

Finally, Google hasn’t been idle in the education sector. Google’s official blog introduced a new feature of Gemini—Interactive Images.

Learning science tells us that passively looking at charts is far less effective than active engagement. Gemini now allows users to directly click on various parts of an image when learning complex concepts (such as the cell structure in biology or the digestive system).

• Click to Explore: Imagine you are reading an article about the cell nucleus. Previously, you could only read the text description next to it. Now, you can directly click on the nucleus in the image, and Gemini will pop up a detailed definition, explanation, and even let you ask follow-up questions about this specific part.

Although this sounds like a small feature, it breaks the barrier between “text” and “image,” turning static learning materials into a dynamic exploration interface, which is definitely a boon for the student demographic.

Frequently Asked Questions (FAQ)

To help everyone digest this information more quickly, here are a few key questions:

Q1: What specific problem does Anthropic’s Dual Agent Model (Initializer & Coding Agent) solve?

Current AI models (like Claude) often fail tasks when handling long-duration tasks across multiple chat windows because they “forget” previous progress. Anthropic’s solution is to break down the task: the Initializer Agent is responsible for the initial environment setup and planning, while the Coding Agent handles the subsequent incremental development. Combined with Git version control and progress log files, this allows the AI to quickly grasp the situation by reading documents even when “changing shifts” (opening a new conversation), ensuring the project continues to advance without interruption.

Q2: Is there a high barrier for designers who don’t know programming to use Adobe Project Graph?

Although Project Graph uses a “node-based” interface similar to programming, its core purpose is to allow designers to arrange and combine creative tools visually. Its advantage lies in allowing high-level creators to build complex workflows and “encapsulate” them into simple tool interfaces. For general users, they may not need to personally drag wires and connect nodes, but can directly use tools shared by others that are already encapsulated, which actually lowers the barrier to using advanced AI techniques.

Q3: Does Perplexity’s memory feature have privacy concerns?

Any feature involving AI remembering personal preferences and history inevitably comes with privacy considerations. Perplexity emphasizes that this is to provide a more precise, personalized search experience. Users can usually manage these memory preferences in settings. From a practical standpoint, this significantly reduces the trouble of repeatedly providing background information—for example, once the AI remembers your programming language preference or dietary habits, the answers it gives will hit the mark directly.

Q4: What types of images does Google Gemini’s Interactive Images feature support?

Currently, this feature is mainly optimized for academic and educational content, especially charts that are structurally complex and require annotation, such as biological anatomy diagrams, mechanical structure diagrams, etc. Google’s goal is to transform passive reading into active exploration with learning materials through this interactivity (clicking specific image areas to get explanations), thereby improving learning outcomes.