正在加载视频...
视频加载失败
Google just showed how AI engineers are moving from simple "RAG" to "graphs", memory and multimodal agents • 15:32 - setting up the production agent stack • 28:00 - turning disconnected data into a knowledge graph • 41:00 - Graph RAG with semantic + hybrid search • 58:00 -... show more
74,975 次观看 • 3 个月前 •via X (Twitter)
21 条评论

persistent memory makes graph rag genuinely useful

yeah, it's true

wow, a new video from Google. booked

yeah - must watch video

Graphs and memory make agents much more interesting, especially when you realize what cool things you can do with them

did you build your own graphs?

No, of course not, I didn't even have that idea, that's why I liked the article, I want to try it too

@0xMorlex this is the stuff we live for, right? connecting data to create knowledge graphs changes everything. makes me want to rethink how i structure my AI projects.

The jump from RAG to graphs is really about grounding. RAG can retrieve the right passages. A knowledge graph helps the agent understand how those pieces relate, which matters when multi-step reasoning starts moving across disconnected data.

Interesting how fast we're shifting from RAG to memory and graph context. I wonder if MENA underfunded R&D can keep up, or if we'll be stuck importing these tools in 2030.

Context graphs are useful because relationships answer questions that chunks cannot. The trust layer is edge provenance: which source created each relationship, when it changed, and whether the agent may act on it or only cite it.

Context graphs could be the missing piece for smarter agents

The jump from flat RAG to graph-based memory feels like the real unlock for multimodal agents — most of my early agent bugs came from context with no relational structure. Saving the graph RAG timestamp.

Grate share

The section on turning disconnected data into a knowledge graph and then layering Graph RAG + persistent memory is pure gol: Watching them go from basic RAG all the way to multimodal agent graphs with Spanner Graph, hybrid search, and memory that survives across sessions. That alone is already worth the full 94-minute workshop.

Standard RAG worked like a basic search engine, pulling text without holding long-term context. Every new session meant wiping the slate clean. The shift happening now: Knowledge graphs to connect relationships Multimodal inputs beyond simple text Persistent memory across sessions That last piece is where Walrus Memory comes in. We give agents a dedicated storage layer so their memory stays portable and verifiable without resetting between runs.

持久記憶讓圖譜 RAG 真正實用

The multimodal extraction part (58:00) is especially interesting. Curious how well Spanner Graph handles noisy video/image-derived entities in practice compared to pure text graphs.

This workshop shows the real progression clearly. Moving from basic RAG into graphs with persistent memory is where agents stop resetting every session and start compounding actual knowledge.

What I keep examining when I write about Graph Engineering is the boundary between retrieving a matching paragraph and preserving relationships across several updates. Once a later decision depends on those links, plain RAG gets harder to audit. I’ve collected the graph, memory, and verification patterns here:

RAG to graph memory, definitely bookmarking this
