1. Introduction: The "Memory Loss" Problem in Long-Term Projects

When developing a large-scale web app with Genspark, we faced the problem of AI completely forgetting previous instructions. Even when asked to "implement it according to the database schema specified earlier," it would ask back, "Which schema are you referring to?"

The cause of this problem is the limits of the "context window." There is an upper limit to the amount of information an AI can process at once, and in long conversations, old information is automatically forgotten.

In this article, we will introduce three strategies that have proven effective. By implementing these, you can maximize the AI's capabilities even in long-term projects.

2. What is a Context Window?

A context window refers to the range of conversation that an AI can refer to at one time. It's similar to human short-term memory, and its capacity has limits.

2.1 Genspark's Context Window

Genspark's exact context window size has not been disclosed, but based on practical experience, it is estimated that approximately 80-100 messages is the limit. Beyond this, old information is automatically deleted.

2.2 Why Are There Limitations?

• Computational Cost: Processing all conversation history requires vast resources
• Response Speed: Larger context increases processing time
• Accuracy Decrease: Too much information can lead to losing important details

3. Signs of Limitation: 5 Signs Your AI is "Becoming Dumb"

When the context window limit is reached, the AI's response quality visibly degrades. Pay attention to the following signs.

3.1 Sign 1: Forgetting Previous Instructions

"Use the naming convention we decided earlier" → "Which naming convention are you referring to?"

3.2 Sign 2: Making Contradictory Suggestions

It makes suggestions that contradict past decisions, such as "We decided to use PostgreSQL earlier, but let's implement it with MySQL this time."

3.3 Sign 3: Repeating the Same Questions

It repeatedly asks for information already provided, such as "Please tell me the database connection information."

3.4 Sign 4: Decreased Level of Detail

Initial responses were specific, but they gradually become abstract and vague.

3.5 Sign 5: Losing Sight of the Overall Project Picture

It becomes unable to answer basic questions like "Which file should I write this code in?"

4. Strategy 1: Building an External Memory System

Without relying on AI's memory, build an external, persistent memory system.

4.1 Utilizing AI Drive

Genspark's AI Drive is ideal as a central management location for project information.

4.2 Creating a Master Document

Create a project "Master Document" and keep it always up-to-date.

4.3 Information Sharing on a New Chat Screen

When migrating the chat screen, share the link to the master document first.

5. Strategy 2: Regular Chat Screen Migration

Chat screen migration is important to perform proactively.

5.1 Timing of Migration

• Number of Messages: Migrate when 80-100 messages are reached
• Switching Development Phases: e.g., Design → Implementation, Implementation → Testing
• End of Day: Start with a new screen the next day
• Signs of Quality Degradation: Migrate immediately if any of the 5 signs mentioned earlier are observed

5.2 3-Step Migration Procedure

5.3 Post-Migration Verification

Ask simple test questions on the new screen to confirm that the context has been correctly transferred.

6. Strategy 3: Information Compression and Summary Utilization

To efficiently use the context window, compress information.

6.1 Reducing Unnecessary Information

• Avoid Redundant Explanations: Aim for concise instructions
• Remove Past Failure Stories: Details of resolved errors are unnecessary
• Diffs instead of Full Code: Share only changed parts

6.2 Regular Summary Creation

Every 20-30 messages, ask "Summarize the content so far" and save it to AI Drive.

6.3 Structured Information Sharing

Organize and share information hierarchically.

7. Summary: Context Management is the Developer's Responsibility

The limitations of the context window are an AI technical constraint and cannot be completely avoided. However, with appropriate strategies, their impact can be minimized.

Context management is the developer's responsibility. Instead of entrusting everything to AI, humans can maximize AI's capabilities even in long-term projects by appropriately organizing and managing information.

Reference Links

• OpenAI GPT-4 Technical Report
• Anthropic: 100K Context Windows
• arXiv: Long Context Language Models
• Qiita - LLM Articles
• Zenn - ChatGPT Topics