NEW Topological RAG for Temporal AI Memory - CRAZY! 😆

Añadido: 2026-05-19

[00:00:00]Hello community. So great that you are back. You may say this is the latest in EI memory for artificial intelligence.

[00:00:07]Absolutely. We do have here AI memory but now with the additional complexity here in time because what we want we want a temporal semantic structure here.

[00:00:19]And if we have a query here, we build now fragments here and those fragments interact at the same time with a temporal tree dynamic that we're going to build. And this is interwoven on sub queries with a knowledge graph. You see here the activation of the node of the knowledge graph that are relevant for a particular time slot here in our temporal tree structure. And then we bring it all together with a tiny little bit of mathematics. And this is here really the very latest in AI memory development. So ready set go. This is a beautiful new study by the Chinese University of Hong Kong Shenzen and the Huawei cloud computing technology May 15, 2026. Hm. A novel memory mechanism for evolving and retrieving agent memory via a hybrid structure. So they say we have a problem now that cannot only we will build something a novel memory structure that cannot only effectively model the evolution of the agent memory over a long period of time and this is really here about long time periods. You know really robotic structures that you have a conversation about 10 20 minutes but they also have build a structure that is easy to retrieve. So this is now a rack system for memory. Let's have a look at this. So how we build it? Today I also showed you we built here a temporal tree structure that we will interwo we we will inter interact here a little bit the temporal and the semantic complexity in a tree structure that allows you the short-term memory data to evolve progressively into the long-term memory data where the later provides the summarized information about the former and in the same time we have to build here a knowledge graph. So simultaneously we are constructing a knowledge graph to capture here the relationship the edges between the entities here that are living in our memory structure. So if you have a feeling that your memory becomes a little bit more complex a little bit more dynamic and a little bit more mathematically oriented absolutely AI is science and this paper shows it more than any other paper. So it offers an effective memory retrieval approach.

[00:02:39]This is the bonus the cherry on top of the cake by exploiting here this particular hybrid structure that we're going to build of the tree for the time and the graph for the relational structures. So let's have a look what we have up until now. All the different methods you have here. They have an explanation how do they work. They are vectorbased. Yes. They have a memory evolution. They have multihop reasoning capabilities. Yes. No. And you see the very last line here is our new system HM. It's a hybrid of a tree and a graph structure. Structure based. They're gentic vector-based. Has a memory evolution and has multihop reasoning for long reasoning traces for long conversation. we are in here for this long problem here where all the other AI models just crash and crumble and now they say hey we found a solution. So let's have a look. So this is kind of a memory evolution here. This denotes you the temporal window based consolidation from the short-term memory to the long-term memory summaries that are created by OI. While the multihop reasoning denotes here an entity or relation level traversal across the graph to different memory fragments that are also defined and harbored here by our EI system.

[00:03:58]So careful this I saw at the beginning if you just combine the indexes this is not something new now that I want to show you in a video but this is not it is not a tree index that you just append to a graph index. It is really that we do have a dynamic coupling of the temporal semantic memory evolution that is now based with an entity centered multihop reasoning capability of the memory.

[00:04:23]So you might say hey now it gets a little bit interesting no because we have skills and skills. So we have here reasoning and interwoven skill dependencies and now we go for the memory. It is not anymore a simple memory markdown fell.

[00:04:40]So how do we build this? How can we do the coupling? The tree structure of HM organizes the memory data both temporarily and semantically where each tree I showed you this in the video.

[00:04:52]Each tree note retains here the memory information regarding a specific semantic topic or even a subtopic but this is defined within a predefined time window. Let's say this is my conversation with the eye on Monday or for a robot. This is the interaction that is happening here in the time the robot is I don't know positioned in a kitchen. So we have semantic and time here together. So each leaf node of our temporal tree structure stores now what it is an event of the agent's original memory fragment a substructure with reduced complexity containing now a semantic topic. Let's say a message in a conversation. Let's say maybe it's physics or whatever generated at a particular timestamp while the upper level nodes store now the memory summaries of some fine grained semantic topics in their lower levels. This means converging their respective time windows.

[00:05:53]So this is a little bit disappointing because here we just have a memory compactification. No, and we have summaries. We have all the problems that we have with AI summary generation. But okay, let's go with this. This is what the orers here go. Now to enable a memory evolution, it performs now a temporal and semantic consolidation. And this is yet beautiful because think about we are working with two different topological elements. On the one side we have a tree structure and on the other side we have a knowledge graph. So how do they come together? So we are here in the tree structure. So today is given two tree nodes whose time windows are very close and you see immediately what we are going to go in the same level if the semantic similarity between the memory data exceeds here a predefined threshold. Now the semantic similarity you know from our vector store this is nothing else than the cosine similarity and exceeds a predefined threshold. This is an absolute value that I say okay if epsilon is greater than whatever then I would consider that this is here really the sematic similarity is high and I have to take this into account.

[00:07:00]This means then the two tree nodes could share here the same parent node whose memory summary now of the parent node preserves now kind of hopefully the consolidated information of these two nodes. So you see exactly we have a hierarchical memory building. Now a lot of information of the very fine nitty-gritty detailed information can get lost in this way. Yeah. But let's see how far we come with this particular idea. So we have here the temporal tree and it is also a cosine similarity semantic closeness tree structure that allows not a short-term memory to evolve progressively into a long-term memory.

[00:07:42]But we have both we have the complete time development and this is what why we need to form. We couldn't do this in a graph structure but we can do some beautiful things in graph theory. A graph structure of HMAP maintains now the knowledge graph of the entities and their relations their edges that are extracted here from the memory data and normally not the old memory data but we will go with memory fragments here that we split up the memory in some chunks like you know maybe here from vector uh from the classical rag here in the vector stores here we have here memory data effectively recording the entity centered information beyond the temporal order and capturing ing now multihop relationship between entities that live on different memory fragments. So this is what we need to graph for. We have different memory fragments here in different topological subspaces and now we need this for our complex reasoning trace. If we have multi-hop relationships and this goes now beyond the temporal order and now we have to bring them together because each topological element or each topological subspace has some benefit to offer to us. The tree the time and the knowledge graph simply the relational density here.

[00:09:02]So let's use both the tree and the graph structures complement each other and this hybrid structure overcomes here the issue on relying on a single index that is the current prevailing system here and the existing research on existing publications. So very nice. We have something completely new. Not a single merging, not a single uh index compactification.

[00:09:26]But we really use here the how say how the beauty of the topology for the tree and the graph for their structures.

[00:09:37]And what we is here the bonus as I told you includes an effective retrieval method. So how is this in the video you saw that the the lightning bolt that comes down from the query it it dissipated in multiple channels. So same idea here no give me a query Q it's first first step here decomposes the complexity of a high complex query into lower complexity subqueries you know we classical do this with almost everything and then we generate a retrieval workflow for each subquery.

[00:10:11]Okay, split it up, make it little chunks that we can work with that we have the complexity engines to deal with their particular task and then define the workflow for each one of them. So then for each subquery it locates some original memory fragments and multihop relevant entities in the graphs. So this is now again a kind of a search process where we need the graph structure for.

[00:10:37]Yeah. Now could theoretically go something wrong or we do not catch all the important um entities in the graph?

[00:10:47]Yes, of course. But there's also a high probability working here with graph theory and with mathematics that we get here I don't know let's say the most interesting other nodes the other sub networks here on the graph that might carry here additional information. You will see this in the result how this increases the performance compared to the classical methods. Okay.

[00:11:12]And afterwards it searches here the relevant evidence from the tree in a bottom up manner. This is now that we jump over to the tree which is then used for completing here the rack process for the memory the memory fragments and here the complete compactification here to feed all this information just from the memory into the context window of our LLM. I mean think about this. This is amazing.

[00:11:41]If we have our query, everything goes down diverges. We have topological structure evolving here between the tree structure and a graph structure in subn networks that are just crazy, just amazing.

[00:11:57]But let's do this. Let's see what happens if we go with this. So just to make sure where we are, we do have the classical rag pro process here from the vector store you know I don't know five six seven years ago now where we have facts new data new knowledge here because yeah the LLM needed new knowledge from what happened yesterday then about two weeks ago I started to show you the rack process here if we have 10,000 skills on a whatever skill bank here on the internet and Now you have to search for particular skills that are optimal that you need.

[00:12:35]Hopefully the LLM knows that you need the skill number 12, 24 and 128 for your particular job. So then you have the retrieve and augment process here from the skill databases. And today now we have a new rack process for the pure memory because you don't want to start your EI fresh from a blank sheet of paper. You want to have the conversation that you had in the last 3 months with your AI system. So you need here the complete complexity of all your discussion in the last 3 months that is now inscribed in the memory but in a way that it uses not up all the available token in the context window.

[00:13:19]Think about this. This is here our little context window. I thought about I imagined here it is a horizontal bar but here you have Gemini build me here this image. It's a little box. Okay, let's go with this. So what we have? We have our memory and MD file. Yeah. And then we have all our instruction and we have here from the system ROM to the user whatever. Then we have let's say 10 skills that we downloaded here the markdown files. Then we have a process where we build from the skills adaptive skills. We have some self-generating skills optimized for your particular job. So they stuck here on the memory and the foundation and then you have a conversation and then you have further instruction and then you have further sequences that you provide. Then you have few short examples in the context window and my goodness no it fills up like crazy and 100% capacity full and we have not even started to harmonize the representation of the memory with the complexity of the skills and I showed you in one of my last video that if you find here a coherent representation in a mathematical space where those two work beautifully together you can really compactify an excel accelerate here the reasoning process but more about this later coming back to this paper of today HMM.

[00:14:42]So here you see it. We have here a hybrid structure that is here if you want our knowledge network here our knowledge graph and then we have here three structures building up and we have here the time component going here from the left side to the right side and yeah this is a hybrid structure. So let's see we have the entities here in our knowledge graph. Beautiful. Then memory summary is the exact opposite here. Okay here we have the memory summary. We understand here everything that happened in the last 3 months from our conversation with your personal AI system. Then we have a memory event you see here those here is the particular time slot. Then we have the memory fragment that is stored in particular time slot here and this constitutes if you want the short-term memory and then over aggregation and summarization and compactification we built here over our temporal axis here the long-term memory. Beautiful.

[00:15:41]As I told you, not a simple index add-on, a simple um construction here, but they really have here a fine grained way how to build this up. So, they give us here an example. We have a user here, beautiful and a lamp and lamb dishes here and whatever and lamb stew. Okay, so the user says, I don't like lambs, so I order chicken instead. And the AI says then I will avoid suggesting any further lamb dishes. Beautiful. So we have now our if you want fragment use it on like a lamb user order chicken rice over lamp and whatever. So you see you just build this up this complexity over the time over the relationship that you have in your knowledge graph and over the fragment and the summarization here over the memory fragment and the short-term memory elements. There is nothing complex about this in a mathematical level. However, you will see the training this we have a lot of predefined hyperparameter. So we need a lot of training to have here the complexity satisfied of the knowledge graph and the complexity of the temporal tree structure.

[00:16:54]Okay. So officially we have the offline indexing and you know this already we talked about this. Now this is the tree construct. So we build a temporal semantic tree in an incremental manner.

[00:17:06]Assume that the tree has L levels where the leaf nodes are at the first level and each level has two hyperparameters alpha level and beta level where alpha level is the similarity threshold between the node and its child node. So you see we have here the linguistic complexity and the linguistic similarity of our expression and betal is simply the size of the time window that we are working here. time interval that we are recording here and we are calculating here and at each level L the newly inserted node is now assigned to a corresponding temporal window. So this is the way to do it with the size B2L according here to the time stamp and only the existing nodes within the same temporal window and this is important are considered now as candidates for the semantic consolidation. So this is the way to do it. So this means within this temporal window now we compute pair wise semantic similarities between the newly inserted nodes and the existing lower level nodes. So this is a very simple approach to handle this and now comes the interesting part because you know at the same time in parallel we build the graph incrementally builds here the knowledge graph. First it extract here the entities the places the persons all the relation between all the objects here from each original memory fragment f that has been defined here at this level I second the extracted entities are normalized and resolved into the entity nodes this is the classical um methodology that you know through here whatever text normalization lamatization token overlap fuzzy string matching yeah I have a lot of videos here years ago on this and if a resolved entity exactly matches an existing entity node and their types are compatible, it is simply now merged into the node. You don't want to have a dduplication going on in your graph. Otherwise, if it's new, if it's brand new, if it's interesting, new knowledge, it's kept as a new entity node and some associated edges may be inserted and defined and have some additional information. and we build up a complex topology given here the complexity of the domain that we're working with. If you're working let's say in theoretical physics it might be a little bit more complex. It might be a little bit not so easy to build if you work here just in social media and you sort here the I don't know the food images of yesterday.

[00:19:34]Okay, this was the offline indexing and now as I told you we have the bonus the jerry on the cake the online retrieval.

[00:19:42]This is now if you wanted to rag for memory in my very simple simplification in my simple words. So given here query Q hm identifies relevant memory evidence by searching now over this hybrid structure. So this is rag for memory we have the agentic rack of workflow over the if you want memory manifold.

[00:20:03]Now the orers go with here a simplified version. They go here with a retrieval planning. Okay. Then the evidence retrieval and the generation process. So let's have a look. Now planning is clear. Now llm reasoner breaks here the complexity human query my query into dependent subqueries here from K equal 1 to capital K. Crucially it classifies here the required memory scope for each subquery either and this is here the simplest you can think of. No short. So evaluate only leave nodes long evaluate only upper tier abstracted summaries or you have mixed. So we have a classification and now we have the multihop graph expansion. So seed entities are now extracted from our QK and matched here in the knowledge graph and the system traverses now over all the edges to retrieve now a particular semantic specifically linguistic relevant subgraph of topologically relevant entities.

[00:21:06]There's nothing new. We all know how to do this but the combination is absolutely fascinating. And then we have the bottomup tree search. No. So we have now retrieved the entities from the graph. They are mapped back to the source fragments at the leaves of the tree and using this as anchors the algorithm searches now upwards through the temporal semantic tree constrained here by the short long scope to extract here the optimal mix of raw effects and highlevel summaries. And hopefully, fingers crossed, the learning process of this AI system was so brilliant from your training pre-training data and after training data that this system really got the optimal mix of the raw effects of the knowledge and of the highlevel summaries of your memory complexity that you built up over the last three months.

[00:22:01]Tiny tiny little bit of mathematics.

[00:22:03]Unfortunately, how you can build this, how you can do here your objective function, how you can do the training of this system. Now, the simplest way honestly the simplest way you can do it is here you just add here elements factors. So once the algorithm got us here candidate set of memory evidence, we have now what we think we can put into the context window of our LLM. This is what the rack comes back here from the wilderness and tell us look what I found. So it must now rank and filter those chunks of memory elements into a final context window for the generation for the generation LLM.

[00:22:44]So it scores now each piece of memory against the particular subquery that from my master query that I defined as a human at the inference time t using now an objective function. And this objective function has now three elements. And yeah, we have a theta 1, a 2, a three. These are simply hyperparameters that hopefully will be learned and will be optimized in the learning process here. But what is really interesting is what are the elements that they add up here. So the first is the semantic similarity. And you might say of course no this is the standard coen similarity between the dense embedding representation of the memory and the query. This is what we do with the normal rack when we have an incoming query and we have here some vector representation of I don't know whatever it is 10,000 books and then we just go for a co similarity between the embedding representation great this is s this is simple what is t and what is r now t turns out the temporal relevance I told you we're looking here for the temporal development here of the memory dynamics now this is rather complex here. So if the query Q subquery QK contains a timebased constraint, it is mapped to a time interval L subK. The candidate memory M spines now the time intervals L subm on temporal relevance optimizes both the discrete intersection and the continuous translation between the intervals with this beautiful mathematical formula. And we have epsilon here that prohibits here something mathematically to happen that we don't want here. Beautiful. We have a formula for this. Great.

[00:24:28]What else is missing now? Now missing is that you don't want to have noise that now really disturbs your system. So all the irrelevant noise mmatmatically should fade out here after retrieval pipeline. But at the same point you want that the most important information in the memory complexity are now reinforced and are really brought back into the context of internet.

[00:24:51]So you also want highly reinforced abstraction that they remain highly ranked here in the vector space. So they are sure to make it here into your prompt. So they call it the order called this here the memory robustness or and this function here if you want models here the memory retention using a dynamically modulating forgetting curve.

[00:25:14]Now forget about it. There are so many possibilities how to achieve this. You have here a mathematically a function that if you want penalizes here the retrieval here of unreinforced memory.

[00:25:26]So you have exactly what we set out to do. The irrelevant noise will mathematically fade out and highly reinforced abstraction will remain dominant and transferred here to the context meter of our LLM. Beautiful.

[00:25:42]Now the artist give us here their M configuration. I always like this. Now, the first I was a little bit disappointed because they go with a GPD4 omni mini and a GPD 4.1 mini as a backbone LLMs which I thought okay but then I thought okay never mind this is a level where we have our tiny local I don't know 3B models here so let's go with this and see if this methodology on this simpler GPD4 omni models really provide an increase in the performance of those LLMs And they say for the semantic retrieval and the evidence reranking the default configuration uses here the embedding here from Q13 and the Q13 reanker here the 4B. Okay, you can go with any espert system that you prefer. Maybe I have built for the physics my own embeddings and my own ranker and reranker but as you like and they have a lighter configuration here with the embedding and the reranker not the 4 billion but the uh 0.6 six billion structure whatever they go here with eight Nvidia or takes a 5000 GPUs each with 24 GB of RAM. So this is here you can rent this in the cloud here. This is not so expensive. I think currently the price for a cloud A5000 is about $2 an hour.

[00:27:04]So eight times you have an idea benchmarks. Let's have a look at this.

[00:27:10]So on what do you evaluate this because this is not interesting. uh this is now here some long-term agent memory benchmark. You really want to have a long ongoing conversation or complexity here in a robotic task or whatever. And they decided to go here with two standard locomo and long memory evaluation. And then the interesting thing, the closer to the reality is real talk.

[00:27:36]But okay, so we have 1,500 question whatever here we have 500 question standard. Beautiful. Let's do this.

[00:27:43]So yeah, GBD4 Omni Mini is it hurts but okay. So whatever you imagine, imagine here a tiny little three billion local LLM. You know I think this is about the same performance that you see how it optimizes and you go here from the old methods here like me zero or me 3 m OS zap here and you see yes it really brings an improvement and you see for H memory here almost here always here the best performance here on the complete line and if you go with GBD 4.1 mini you'll see okay there's a little bit of a fluctuation but also you have the bold numbers here on Hm. So for the benchmark of locom beautiful it works here with let's say currently tiny mouse and it really brings here an improvement.

[00:28:35]If you look at the overall accum um accuracy here gives you a feeling where we are. We go from 60% to 70% to 77% to 85% and now here and imagine here this is for Omni Mini we are at 88%. So yeah, this really kind of seems to work now.

[00:28:56]And here we have a GBD 4.1 mini. We go here best was 90. Now we have 93 almost the other benchmark here. Okay, you got it. What is really interesting is here real talk the the close to the real situation. And if you see here the jump, yeah, it is still there. But yeah, you see accuracy overall we go from 75 from 76% here to 78%. Okay, there is a little tiny improvement. But do you know the complexity? If we have a a small dimensional space for this, it's okay.

[00:29:38]So this means if we have to build a graph with just I don't know thousand 10,000 nodes, it's okay. But imagine you have to build here topological elements or a knowledge graph with millions tens of millions of node and then you have to build here the temporal trees also with a complexity. I think the computational cost I have not done this but the computational cost I can imagine will be quite heavy. So all of this just to go here two percentage points. So this is interesting but also think about what we have here to do our context window of all the alm it is filling up before we even start a conversation which should be here in this um face bundles but yeah uh Gemini decided to make here something else. Okay whatever. Yeah, as I told you, we have all the memory, all the system instruction, all the different skills that you need, all the coding, all the language, all the reasoning complexities. No, and then the interval from dependency, the rag for the skills, and now we have a rag for the memory complexity.

[00:30:49]Then we have the rag for databases, for new facts, for knowledge. And then whatever a few short example you provide here and the context window fills up so fast and then you have maybe a conversation of 1 2 3 4 5 days no and then you have to have somehow a memory compactification because you want to go on over months in your work with your work EI if you want. So I think we have not solved it at all. We just added complexity. We added what we know how to do. We know how to do a knowledge graph.

[00:31:26]We know how to build tree structures.

[00:31:28]And now we've used them together. Made it a lot more complex for highdimensional data. It gets extreme intensive to calculate all of this. And I think this will fill up a context window even I don't know of 250k or up maybe up to 1 million token here rather fast before we even can start to work with this AI and we can provide here the complexity and all the instrument and here the experimental data into the context window of the LLM. So I'm not sure. I think this shows us that it's not the solution. And you said what is the way out? And honestly I don't know.

[00:32:06]But I have some ideas and there are some research going on globally in China in Japan in Korea all over the world in the US even in Europe. You're not going to believe this where people are kind of aware of this complexity and they are trying to find a new solution to this problem. Yeah. Because have I told you that all of this what we have here we even have not started to harmonize the skills with the memory complexity on a topological element like a manifold structure as I have shown you in my last three videos. No, this is still missing.

[00:32:40]We have not even touched it that we have to optimize these elements together to bring in a coherent base element. No.

[00:32:48]Yeah, I think this would be all a topic here for one of the next video. I hope you enjoyed it. I hope you had some new information. Would be great to see you in my next video.