"RAG Is Dead": The Industry Pivots to Agentic, Living Knowledge

"RAG Is Dead": The Industry Pivots to Agentic, Living Knowledge

LlamaIndex's blunt "RAG is dead, long live agentic retrieval" has become a rallying cry following discussions at AWS re:Invent 2025 [1]. The argument: naive vector-search RAG can't support agents that need multi-step reasoning, dynamic namespace selection, and knowledge bases that update themselves. NVIDIA's engineering team frames it similarly — agents need dynamic knowledge, not frozen snapshots [2].

The numbers being cited are significant. Gartner-sourced projections say semantic, agentic approaches could deliver up to 80% accuracy gains and 60% cost reductions by 2027 compared to static retrieval pipelines [3]. Towards Data Science's "context engineering" framing pushes the same conclusion from a different angle: semantic layers, not bigger context windows, are what make agents reliable [3].

X commentary is explicitly linking this to the Karpathy wiki trend — both point toward the same destination: durable, structured, self-correcting knowledge rather than ephemeral retrieval. For anyone building on meeting data, this is the architecture bar to clear.

Claude Agents Turn Granola Transcripts Into Org Intelligence

Granola's background transcription — no bot joining your call — is proving to be fertile ground for a new category of "AI chief of staff" agents [2]. One widely shared build walks through wiring Claude Code to Granola transcripts via MCP, alongside strategy docs and email, to generate dynamic org charts, relationship scores, and proactive advice — including warnings before you send a poorly-timed message to a stakeholder [1][3].

This isn't a one-off hack. It reflects a broader shift where meeting transcripts stop being a record of what happened and start functioning as a live input to organizational reasoning. Examples circulating on X show automated executive briefings assembled entirely from historical meeting data, with agents flagging who to build alliances with based on patterns across dozens of calls [3].

The mechanics matter here: it's the connector access (MCP) plus full transcript fidelity that makes this possible. Tools that only offer summaries, not raw searchable transcripts, can't support this kind of agent.

Meetily Bets on Local-First Privacy With Open-Source Diarization

Meetily launched as a fully local meeting assistant for Mac and Windows, built in Rust with Parakeet/Whisper for transcription (claimed 4x real-time speed) and Ollama for on-device summarization [1][2]. No audio leaves the machine. Version 0.3.0 added audio import and retranscription, and the project's GitHub presence — 359.5K stars per the project's own site — suggests real developer traction [2][3].

The pitch is squarely against the cloud-tool fatigue narrative: Meetily's own materials position it as an alternative to Otter, and X discussion around the launch leans into complaints about running Granola, Otter, and Fireflies simultaneously with overlapping subscriptions [3]. Speaker diarization out of the box is the headline feature that puts it in more direct competition with commercial tools.

It's a useful signal, even if privacy-first local tools trade off the compounding knowledge-graph capabilities that cloud-connected systems (like the Claude/Granola stack above) can offer.

What This Means For Your Meetings

Today's stories all converge on one idea: meeting data is only as valuable as the memory system wrapped around it. Karpathy's LLM Wiki and the "RAG is dead" consensus are really the same argument applied to two different domains — static retrieval, whether over documents or meeting transcripts, is being replaced by living, cross-referenced knowledge that gets smarter with every addition. A transcript from March should inform how an agent interprets a transcript from July. That's not a search problem; it's a knowledge-graph problem.

The Granola-Claude experiments make this concrete for professionals: meeting history becomes an input to judgment, not just a record of decisions. Org charts, relationship intelligence, and proactive nudges only emerge when an agent can reason across dozens of meetings simultaneously — exactly the pattern Proudfrog was built around, connecting speaker identity, topics, and decisions into a persistent knowledge graph rather than a folder of transcripts. Meetily's local-first approach is a reminder that not everyone wants their meeting data leaving the building — but it also shows the tradeoff: privacy-maximalist tools today are optimized for transcription quality, not for compounding organizational memory across meetings.

The practical lesson for anyone drowning in Zoom calls: the tool that wins isn't the one with the best transcript — it's the one that remembers what the transcript meant six months later, and can tell you why it matters again today.

Key takeaway: Static transcripts are becoming commodity; the real value — and the real competitive battleground — is in the living, cross-meeting knowledge graph that lets you retrieve not just what was said, but what it means now.

Sources

1. https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
2. https://aimaker.substack.com/p/llm-wiki-obsidian-knowledge-base-andrej-karphaty
3. https://medium.com/@roanmonteiro/building-a-complete-personal-harness-llm-wiki-developers-second-brain-in-obsidian-d7b61c7398ff
4. https://www.llamaindex.ai/blog/rag-is-dead-long-live-agentic-retrieval
5. https://developer.nvidia.com/blog/traditional-rag-vs-agentic-rag-why-ai-agents-need-dynamic-knowledge-to-get-smarter/
6. https://towardsdatascience.com/beyond-rag/
7. https://techysurgeon.substack.com/p/how-i-built-an-automated-executive
8. https://www.granola.ai/
9. https://mcpmarket.com/tools/skills/granola-meeting-notes
10. https://github.com/Zackriya-Solutions/meetily
11. https://meetily.ai/
12. https://meetily.ai/blog/self-hosted-meeting-note-taker-guide-2026