Qwen: AI Memory "NapMem" as a Dynamic Action Space

This is insane... I just built this for Hermes... independently of knowing about this.

best creator about ai on the net. with out a doubt.

Thank you for sharing!! I'm currently using Raptor, which is very fast but often irrelevant. NapMem is quite expensive in latency, but in exchange, you make sure memories are relevant.

Memory-as-action-space raises the thing your "Amnesia, Malware and Collapse" video already named: the agent is now taking actions based on memory it has no way to verify. If a memory can be navigated, it can be poisoned. We went the other way and write every atom into a SHA-256 Merkle tree, so each read carries a proof - a tampered memory fails the check instead of being silently trusted. Compact RFC 6962 encoding keeps the context cost sane: 2,580 tokens vs 4,102 for the verbose form. Genuinely curious whether integrity is out of scope for NapMem by design, or just assumed away by the benchmark setting. Feels like it stops being ignorable the moment memory is shared between agents.

IMHO memory is not about recall but about generation-on-recall based on semantic anchors from different anchor types(temporal, spatial, semantic), for us the source of truth is the non-stop inflow of data from the outside world, in a "rag" setup the bottom layer should be chunks and exact documents or pointers to real bytes stored somewhere, but in a common agent "assistant" use-case most queries should be answerable without reaching to the exact deterministic layer at all (or maybe use it only for internal grounding - will see once I get some real-usage data) - layered architecture with "zoomable" S2-like recall is the right "shape" or direction

At some point it will be more effective to merge the active memory processes into the actual neural network into an active learning structure, similar to what the Future AI Society organization describes. Also, the industry keeps trying to predefine things, when the neural network itself should have rules to be able to add, delete, edit, its own things. For example instead of predefining (User Profile, Topic Tracks, Memory Records, Raw Conversations) why not create rules and processes that allow the neural network to create those four or more categories, since those four categories are not going to be optimal in all situations and industries, but if the neural network can has rules, tools, guidance so that it can determine and measure what would work best it can self assemble what is optimal depending on the situation. The more the neural network can define and adapt its own architecture the better.

Ah the classic..."psoidocode".