想问问你们什么时候能支持记忆的重要性评分、去重、命中追踪、访问评率，还有衰减机制？

[rexia01]

如题目，谢谢。

[freshpomelo]

Could you please provide more details about your application scenario? When will you use these memory features?

[rexia01]

Could you please provide more details about your application scenario? When will you use these memory features?

I think all of these are related to the accuracy of memory retrieval, right? Aren't these the most basic things of memory system? You can refer to mem0.

[freshpomelo]

Appreciate your insights. In my view, a truly robust memory system should offer a seamless end-to-end query experience, minimizing the configuration burden on agent developers. And still some features is in our planning, stay tuned for updates.

[rexia01]

Appreciate your insights. In my view, a truly robust memory system should offer a seamless end-to-end query experience, minimizing the configuration burden on agent developers. And still some features is in our planning, stay tuned for updates.

I completely understand what you're saying, but my main question is about how to increase the accuracy of memory retrieval in your memory system, especially given the ever-expanding memory base. Is it about increasing the weight of memories or using decay mechanisms? Like the human brain, it can't remember everything.

We have now integrated with your product, but we have found that your code does not have a mechanism for this. How can we improve the accuracy of our product in terms of memory retrieval?

[BaoBaoGitHub]

Appreciate your insights. In my view, a truly robust memory system should offer a seamless end-to-end query experience, minimizing the configuration burden on agent developers. And still some features is in our planning, stay tuned for updates.

Here’s why I think these mechanisms matter.

conflicts
User preferences change: “I hate Lee” yesterday → “Actually I don’t” today.
It’d help if the system can update/override old or wrong memories.

Importance + time decay
Not all memories are equal: yesterday’s lunch should rank above last year’s lunch, but a big event last year (anniversary) may rank above a small thing last week.
So we’d like importance scoring + time decay + usage frequency (memories used often get higher weight).

merge
If users repeat the same topic a lot, creating many duplicate MemCells, resulting in redundancy and slower retrieval.

[ZhenhangTung]

Appreciate your insights. In my view, a truly robust memory system should offer a seamless end-to-end query experience, minimizing the configuration burden on agent developers. And still some features is in our planning, stay tuned for updates.

Here’s why I think these mechanisms matter.

conflicts User preferences change: “I hate Lee” yesterday → “Actually I don’t” today. It’d help if the system can update/override old or wrong memories.

Importance + time decay Not all memories are equal: yesterday’s lunch should rank above last year’s lunch, but a big event last year (anniversary) may rank above a small thing last week. So we’d like importance scoring + time decay + usage frequency (memories used often get higher weight).

merge If users repeat the same topic a lot, creating many duplicate MemCells, resulting in redundancy and slower retrieval.

@BaoBaoGitHub Do you have any real-world agent use cases you can share where conflicts / time-decay / merging actually made a difference?

[ZhenhangTung]

Appreciate your insights. In my view, a truly robust memory system should offer a seamless end-to-end query experience, minimizing the configuration burden on agent developers. And still some features is in our planning, stay tuned for updates.

I completely understand what you're saying, but my main question is about how to increase the accuracy of memory retrieval in your memory system, especially given the ever-expanding memory base. Is it about increasing the weight of memories or using decay mechanisms? Like the human brain, it can't remember everything.

We have now integrated with your product, but we have found that your code does not have a mechanism for this. How can we improve the accuracy of our product in terms of memory retrieval?

@rexia-rx That’s really interesting to hear. As an early adopter, would you mind sharing a bit more context about your use case?
For example: what kind of agent you’re building, what the memory represents, and where retrieval accuracy becomes a problem for you.

[BaoBaoGitHub]

Appreciate your insights. In my view, a truly robust memory system should offer a seamless end-to-end query experience, minimizing the configuration burden on agent developers. And still some features is in our planning, stay tuned for updates.

Here’s why I think these mechanisms matter.
conflicts User preferences change: “I hate Lee” yesterday → “Actually I don’t” today. It’d help if the system can update/override old or wrong memories.
Importance + time decay Not all memories are equal: yesterday’s lunch should rank above last year’s lunch, but a big event last year (anniversary) may rank above a small thing last week. So we’d like importance scoring + time decay + usage frequency (memories used often get higher weight).
merge If users repeat the same topic a lot, creating many duplicate MemCells, resulting in redundancy and slower retrieval.

@BaoBaoGitHub Do you have any real-world agent use cases you can share where conflicts / time-decay / merging actually made a difference?

We’re planning an AI companion toy and want to add a memory module. The project is still in the early planning stage.

One conflict scenario we’re considering is changing user preferences. For example, a user may say they like a certain person at first, but later dislike that person because of something that happened. With the current MemCell approach, retrieval might bring back both “the user likes X” and “the user dislikes X”, which could confuse the model.

We also think time-decayed memory is important. For instance, the user said in January 2025 that they ate cake and thought it was delicious. In January 2026, they ate strawberries and thought they were delicious. Later, the user says: “The tasty thing I ate two days ago—I want to eat it again.” The current memory retrieval might return both “cake” and “strawberries”, but obviously strawberries should have higher priority than cake because it’s more recent.

Overall, in my understanding, memory needs both initial construction and ongoing maintenance. Maintenance includes time-based importance decay, conflict resolution mechanisms, and memory cleanup.

What do you think?

[ZhenhangTung]

Appreciate your insights. In my view, a truly robust memory system should offer a seamless end-to-end query experience, minimizing the configuration burden on agent developers. And still some features is in our planning, stay tuned for updates.

Here’s why I think these mechanisms matter.
conflicts User preferences change: “I hate Lee” yesterday → “Actually I don’t” today. It’d help if the system can update/override old or wrong memories.
Importance + time decay Not all memories are equal: yesterday’s lunch should rank above last year’s lunch, but a big event last year (anniversary) may rank above a small thing last week. So we’d like importance scoring + time decay + usage frequency (memories used often get higher weight).
merge If users repeat the same topic a lot, creating many duplicate MemCells, resulting in redundancy and slower retrieval.

@BaoBaoGitHub Do you have any real-world agent use cases you can share where conflicts / time-decay / merging actually made a difference?

We’re planning an AI companion toy and want to add a memory module. The project is still in the early planning stage.

One conflict scenario we’re considering is changing user preferences. For example, a user may say they like a certain person at first, but later dislike that person because of something that happened. With the current MemCell approach, retrieval might bring back both “the user likes X” and “the user dislikes X”, which could confuse the model.

We also think time-decayed memory is important. For instance, the user said in January 2025 that they ate cake and thought it was delicious. In January 2026, they ate strawberries and thought they were delicious. Later, the user says: “The tasty thing I ate two days ago—I want to eat it again.” The current memory retrieval might return both “cake” and “strawberries”, but obviously strawberries should have higher priority than cake because it’s more recent.

Overall, in my understanding, memory needs both initial construction and ongoing maintenance. Maintenance includes time-based importance decay, conflict resolution mechanisms, and memory cleanup.

What do you think?

@BaoBaoGitHub I strongly agree with your point that “memory needs both initial construction and ongoing maintenance.”
My current impression is that EverMemOS doesn’t yet provide convenient, first-class support for either of these two aspects.

Using a game analogy: initial construction is a bit like character initialization — defining the character’s baseline traits and starting state. As the user interacts more over time, the character gains experiences, upgrades equipment, and changes behavior. All of that needs to be continuously maintained and updated.

I'm curious, if you’re building an AI companion toy, wouldn’t personalization at the individual toy level be extremely important?

[BaoBaoGitHub]

Appreciate your insights. In my view, a truly robust memory system should offer a seamless end-to-end query experience, minimizing the configuration burden on agent developers. And still some features is in our planning, stay tuned for updates.

Here’s why I think these mechanisms matter.
conflicts User preferences change: “I hate Lee” yesterday → “Actually I don’t” today. It’d help if the system can update/override old or wrong memories.
Importance + time decay Not all memories are equal: yesterday’s lunch should rank above last year’s lunch, but a big event last year (anniversary) may rank above a small thing last week. So we’d like importance scoring + time decay + usage frequency (memories used often get higher weight).
merge If users repeat the same topic a lot, creating many duplicate MemCells, resulting in redundancy and slower retrieval.

@BaoBaoGitHub Do you have any real-world agent use cases you can share where conflicts / time-decay / merging actually made a difference?

We’re planning an AI companion toy and want to add a memory module. The project is still in the early planning stage.
One conflict scenario we’re considering is changing user preferences. For example, a user may say they like a certain person at first, but later dislike that person because of something that happened. With the current MemCell approach, retrieval might bring back both “the user likes X” and “the user dislikes X”, which could confuse the model.
We also think time-decayed memory is important. For instance, the user said in January 2025 that they ate cake and thought it was delicious. In January 2026, they ate strawberries and thought they were delicious. Later, the user says: “The tasty thing I ate two days ago—I want to eat it again.” The current memory retrieval might return both “cake” and “strawberries”, but obviously strawberries should have higher priority than cake because it’s more recent.
Overall, in my understanding, memory needs both initial construction and ongoing maintenance. Maintenance includes time-based importance decay, conflict resolution mechanisms, and memory cleanup.
What do you think?

@BaoBaoGitHub I strongly agree with your point that “memory needs both initial construction and ongoing maintenance.” My current impression is that EverMemOS doesn’t yet provide convenient, first-class support for either of these two aspects.

Using a game analogy: initial construction is a bit like character initialization — defining the character’s baseline traits and starting state. As the user interacts more over time, the character gains experiences, upgrades equipment, and changes behavior. All of that needs to be continuously maintained and updated.

I'm curious, if you’re building an AI companion toy, wouldn’t personalization at the individual toy level be extremely important?

AI toys come with personalized settings, which are used to configure the toy during the initial setup.

However, we also hope the AI toy can build and continuously update its memory based on the user’s chat history, so it can better understand the user over time and provide more thoughtful, long-term companionship.