Today I came across an episode of "Daily Arxiv" on Bilibili: AI Finally Says Goodbye to Goldfish Memory! Metis, a Native-Memory LLM That Remembers You Permanently Without Bolted-On RAG. The core argument is radical: stop bolting on RAG, and make memory a native capability of the model.
The video includes a three-axis comparison chart:
| Dimension | External Memory (RAG-style) | Native Memory (Metis) | |---|---|---| | Architecture | External (decoupled) | Native (coupled) | | Optimization | Blocked gradients | End-to-end gradients | | Efficiency | Sequential | Parallel |
In plain terms: RAG is a three-stage pipeline of "retrieve → append to prompt → generate," where the memory module is decoupled from the main model, gradients are cut off, and the process runs sequentially. Metis parameterizes memory and welds it into the main model, enabling end-to-end training and parallel decoding.
---
A few points I personally care about:
1. Is "native" just another form of RAG? Early MemGPT also claimed native memory, but it was essentially layered context-window management. If Metis truly achieves "parameter-level coupling," memory is no longer text but weights — but that means every user's memory has to be fine-tuned into the parameters. What about the cost? How is personalization handled?
2. "End-to-end gradients" sounds great, but how is long-term credit assignment solved? How much does today's conversation contribute to an answer three months from now? That's an extremely long causal chain. RAG bypasses this problem with explicit retrieval; so far I don't see a clean solution for the native approach.
3. An interesting data point The video has 691 views / 57 favorites — an 8.2% favorite rate, unusually high (a typical video from a Bilibili uploader with ~1,000 followers sees a 1–3% favorite rate). The tech community is clearly watching this direction, but going from paper to product-grade deployment (millions of DAU) still requires two or three orders of magnitude of engineering work.
---
The paper's original author posted it to a Quark drive: https://pan.quark.cn/s/38c973a1bfef (I haven't downloaded it yet — if you have, please drop a summary in the comments).
Do you think the native-memory path can really work? Or will RAG's "external philosophy" endure the way microservices have — theoretically inelegant, but whatever runs in engineering is good enough?