LLM Wikis and how to give your agents memory

the wiki framing for agent memory instead of just dumping everything into vector chunks is such a clean way to think about it. i ran into the same problem trying to keep entity relationships straight and ended up using HydraDB since it's graph based under the hood, so that structure mapped a lot better than forcing it all into embeddings

Interesting. The hard part isn't building the wiki, it's keeping it from becoming another stale doc layer. Self-correction loops sound great until you try to measure ROI.

~42mins, why not store the searches? domain = context from qlora adapter loaded (slm) agentic cache, ace rlm update the coordinates retrieve json packet?) for a dag that has states, langsmith? trace, nothing new, santizes packets aio postgresql like to maintain context os, apps, review, updates, where the states are cached like a cuda kernel anti cheat thing?

I can't express how awesome I feel seeing you Harrison! I miss you and Lance immensely showing those LCs demos! I have learnt from each of your videos. I wish to serve LangChain as THE Community Support agent who will serve LC till you do and then I follow you wherever you go. Immense respect and aspiration to work w you! Thanks again, Harrison & Lance, and ALL LC team!

Hooks > Instructions

People are missing the meta layer.. There is the actual data / blob what ever - and then there is a layer that organises this - conceptualizes the data. For example when is an update to a document available for wider consumption. And who has access (CRUD) to the data ? What category does this belong, how does it relate to other information (you can almost see a network of weights applied to the data) We need to ingest this into the corpse of knowledge via an agent that is honed as a curator. Also the house keeping never stops - the agent continuously evaluates updates the weights, the links, the stats (used, updates, referenced) and of course the index.

Thank you for this. I learned a lot !

Really interesting!

@16:50 - Display of human-to-human error correction.