AI developers are incorporating the 19th-century Ebbinghaus forgetting curve theory to help AI discard unnecessary information while preserving vital memories through intelligent forgetting systems.
Imagine this: every morning, you tell your assistant what you need to do. But what if this assistant tries to remember every single word you’ve ever said, down to things from a year ago? Every time you say, “I’m worried about lunch,” the assistant might chime in, “How about the kimchi stew you had on March 15th last year?” and the conversation would be delayed while they dig up irrelevant information.
A similar concern is growing in the field of Artificial Intelligence (AI). As AI models become smarter and try to remember more information, they often experience slowdowns in processing speed or lose the context of a conversation. To solve this, developers have revived a 140-year-old psychological theory: the ‘Ebbinghaus forgetting curve.’
Why is this problem important?
While we expect AI to act as intelligently as a human, AI’s memory structure is actually quite different. Humans naturally let go of unimportant information, but AI tends to cling to all data every time it receives new input. The problem is that this ‘indiscriminate memory’ makes AI sluggish.
Actual research shows that for every 5 kilobytes (KB) of additional memory data given to an AI agent (an AI performing a specific task), the time it takes to process information and make decisions increases by 1.1 milliseconds (ms)[Source: HackerNoon]. This creates massive bottlenecks in services where hundreds or thousands of users access AI simultaneously. If we expect faster response times from AI, the AI must learn ‘how to discard effectively.’
In Simple Terms: AI ‘Memory Diet’
The Ebbinghaus forgetting curve is a graph that shows how much information a person forgets over time[Source: ELVTR]. Simply put, we forget most information soon after hearing it, but information we recall repeatedly becomes more deeply embedded in our brains.
Developers have transplanted this principle into AI memory management engines[Source: Towards Data Science].
Think of AI’s memory space as a ‘photo album.’ Previous AI tried to keep every photo taken every day. But AI with ‘intelligent forgetting’ is different. Frequently viewed photos (information the user often asks about or considers important) are moved to the front of the album for longer preservation, while blurry photos never looked at (unnecessary information) are automatically sent to the trash over time[Source: Towards Data Science]. This allows AI to always focus on the ‘information needed right now.’
Where are we now?
Experiments based on this theory are active in the field. Open-source projects and memory management tools are changing how AI stores and retrieves memories by applying this ‘forgetting curve’[Source: DEV Community].
However, there is still a long way to go. Some models in early experimental stages made errors by deleting data based solely on word overlap (string matching) instead of grasping the ‘importance’ of the information[Source: Eris dev blog]. When a human speaks vaguely, saying, “the thing I mentioned yesterday,” the AI should be able to grasp the context, but because it applied mechanical deletion criteria, it mistakenly deleted valuable context along with the unimportant data.
Also, the problem of ‘amnesia’—where necessary information disappears midway when multiple AIs exchange data within an AI pipeline (workflow)—remains a significant challenge for developers[Source: linksfor.dev].
What does the future hold?
Moving forward, AI will evolve beyond simply learning from vast amounts of data to learning ‘what information to discard.’ We will see the generalization of methods that assign different ‘memory lifespans’ (TTL, Time-To-Live) to different data, rather than the previous approach of managing memory mainly based on the most recent information[Source: TianPan.co].
For example, a ‘performance debugging task’ the user is working on today might be remembered by the AI only for the day, while ‘user preferences or tastes’ are designed to fade away more slowly over a longer period[Source: TianPan.co]. Once this happens, the AI will understand our style like a long-time assistant, without us having to explain everything anew each time.
MindTickleBytes’ AI Reporter’s Perspective For AI to be smart, it needs the wisdom of knowing ‘what to pretend not to know’ rather than just knowing a lot. It is paradoxical yet fascinating that 140-year-old psychological theory is making the brains of cutting-edge AI lighter and faster. Future AI will compete not on ‘memory’ but on the ‘art of forgetting.’
References
- So this “forgetting curve” did not measure importance at all - Eris dev blog
- I built a forgetting curve for an agent with one user - Hacker News
- Multi-agent AI pipelines lose context at every handoff between agents - linksfor.dev
- Forgetting is not passive at all. It is active. - Foxfire
- German psychologist Hermann Ebbinghaus built a forgetting curve - ELVTR
- Context Windows Forget What Matters — I Built a Usage-Reinforced Decay Engine for AI Agent Memory - Towards Data Science
- Your Memory is a practical open-source MCP server that bakes the Ebbinghaus forgetting curve - DEV Community
- The cost curve exposed its own remedy: trim context every fifty seconds and cap recall at twenty kilobytes - HackerNoon
- This mirrors the Ebbinghaus forgetting curve, where retention decays exponentially - TianPan.co
- Implements Ebbinghaus forgetting-curve retention with usage-based reinforcement - GitHub Topics
- To understand human emotions
- To distinguish between important and unnecessary information to increase efficiency
- To increase storage space infinitely
- Humans remember all information perfectly
- Memory retention of information decreases exponentially over time
- Memory is fixed like a photograph
- Because it consumes a lot of electricity
- Because unnecessary memories slow down the AI's thinking speed
- Because AI lies