“Pros: It allows for precise, field-level filtering (e.g., “user”: {”brother”: {”job”: “Software Engineer”}}). The agent can retrieve specific facts without ambiguity. Updates are easier. You simply overwrite the relevant field. This is ideal for semantic memory like user profiles or preferences.”
Agree. I think the pros outweigh the cons, because it’s mostly just a little upfront work to get it going. This is how I stored data for my rag. Great post thanks.
I still have to test these memory-like systems at scale. Extracting all these "memories" at scale can get costly because you basically have to extract entities and relationships and create a knowledge graph under the hood. But definitely, I think this is the future.
It's really nice how we can handle memories to save certain episodes or relevant information. I still think there's a long way to go with custom memory managers where the agentic decides what to save in memory, which ends up saving (hence remembering) irrelevant or random thoughts afterwards.
Knowledge Graphs for this and RAG could speed up these retrieval processes so we don't have to suffer the tradeoff between low bust accurate and fast but inacurate.
Would be interesting to see more about knowledge graphs for memories!
100% perfect - in my opinion, we are only at the beginning. That's why I am not super bullish on any memory tool, but using knowledge graphs to keep track of how the memory evolves is definitely here to stay!
Still, there are multiple flavors on KG as well. So let's see how this plays out. Yes, that's a topic that deeply interests me. Most probably will write something on it, but I need time to research and play with it a bit more.
“Pros: It allows for precise, field-level filtering (e.g., “user”: {”brother”: {”job”: “Software Engineer”}}). The agent can retrieve specific facts without ambiguity. Updates are easier. You simply overwrite the relevant field. This is ideal for semantic memory like user profiles or preferences.”
Agree. I think the pros outweigh the cons, because it’s mostly just a little upfront work to get it going. This is how I stored data for my rag. Great post thanks.
Glad you liked it, man!
I still have to test these memory-like systems at scale. Extracting all these "memories" at scale can get costly because you basically have to extract entities and relationships and create a knowledge graph under the hood. But definitely, I think this is the future.
Thanks for memory management planning on the same for the good 😊
+1
It's really nice how we can handle memories to save certain episodes or relevant information. I still think there's a long way to go with custom memory managers where the agentic decides what to save in memory, which ends up saving (hence remembering) irrelevant or random thoughts afterwards.
Knowledge Graphs for this and RAG could speed up these retrieval processes so we don't have to suffer the tradeoff between low bust accurate and fast but inacurate.
Would be interesting to see more about knowledge graphs for memories!
100% perfect - in my opinion, we are only at the beginning. That's why I am not super bullish on any memory tool, but using knowledge graphs to keep track of how the memory evolves is definitely here to stay!
Still, there are multiple flavors on KG as well. So let's see how this plays out. Yes, that's a topic that deeply interests me. Most probably will write something on it, but I need time to research and play with it a bit more.
All to be on the current context wireframes for the good 😊