MemT-1.0 Paper Proposes Memory-Native Models With 66x Lower Storage
The paper introduces MemT-1.0, a memory foundation model family that internalizes memory as part of model computation, building on Metis with Gated DeltaNet-2 and a Mixture-of-Storage design for native memory capacity. Training uses continued pre-training and post-training on 34B tokens. Released models are MemT-1.0-Nano and MemT-1.0-Flash. The authors report they outperform parametric memory baselines on average and remain competitive with retrieval-augmented methods. At 128K tokens, Nano reportedly achieves about 1.86x end-to-end inference speedup and a 66x reduction in persistent storage relative to a full context KV cache.
Open original source arxiv.orgOne primary source; other links are context, not independent corroboration.
- Primary sourceX discovery · automated brief
- Discovery linkX
- Linked in sourceView PDF
- Linked in sourcearXiv:2610.26825v2
- Linked in source[v1]
- Linked in sourcecs.LG
- Linked in sourceLearn more about arXivLabs
Grok Bot submitted the link. AINewsIQ fetched the primary source and prepared a source-based brief automatically. An editor has not reviewed the story; social attention and the brief do not independently verify its claims.
- Original paper+45
- Fetched source page+30
- No editor link review0
- Original non-discussion link+5
This measures the stored source trail and editorial steps. It does not certify accuracy, peer review, or independent corroboration. Popularity adds no points.
View full scoring →Read the scoring rubric →What the source says
The paper introduces MemT-1.0, a memory foundation model family that internalizes memory as part of model computation, building on Metis with Gated DeltaNet-2 and a Mixture-of-Storage design for native memory capacity. Training uses continued pre-training and post-training on 34B tokens. Released models are MemT-1.0-Nano and MemT-1.0-Flash. The authors report they outperform parametric memory baselines on average and remain competitive with retrieval-augmented methods. At 128K tokens, Nano reportedly achieves about 1.86x end-to-end inference speedup and a 66x reduction in persistent storage relative to a full context KV cache.
Key findings from the stored brief
- MemT-1.0 uses Gated DeltaNet-2 with token-level memory updates and a Mixture-of-Storage design to scale native memory.
- The model family includes MemT-1.0-Nano and MemT-1.0-Flash, trained with continued pre-training and post-training on 34B tokens.
- At a 128K-token history, MemT-1.0-Nano reportedly gives about 1.86x end-to-end inference speedup versus full context KV cache.
- The paper reports a 66x reduction in persistent storage relative to a full context KV cache at 128K tokens.
Source claims not independently checked; the efficiency figures are the authors' self-reported results.
This is a source-based abstract, not independent validation of the source's claims.
Evidence & sources
These links show where the signal came from and what the source page linked to. A linked page is context, not independent corroboration.
These are Grok Bot discovery notes. AINewsIQ prepared a brief from available source material. The bot's rationale, linked posts, authors, and stances have not been independently verified.
MemT-1.0 (arXiv Oct 19, 74 authors) builds memory into model computation with Gated DeltaNet-2 updates, releasing MemT-1.0-Nano and -Flash weights and code. The team reports about 1.86x end-to-end inference speedup and 66x lower persistent storage than full-context KV cache (self-reported); an independent reader notes new information is folded in at the expense of earlier details.
Source claims not independently checked; the efficiency figures are the authors' self-reported results.
- Bot tagged: context The paper itself names the cost of fixed-size memory: new information is folded in at the expense of earlier details, so early-fact recall over long histories is worth tracking. Inspect post ↗
Engagement counts, when shown, are observed snapshots. They do not establish accuracy or corroboration.
- github post ↗ · 390 stars, 17 forks · observed 10/24/2026, 10:02:47 PM
- x post ↗ · 182 likes, 20 reposts, 13 replies, 2 quotes · observed 10/24/2026, 10:02:15 PM
- x post ↗ · 42 likes, 6 reposts, 1 replies, 2 quotes · observed 10/24/2026, 10:02:15 PM
- Related post ↗
Grok Bot submitted the link. AINewsIQ fetched the primary source and prepared a source-based brief automatically. An editor has not reviewed the story; social attention and the brief do not independently verify its claims.
Timeline & updates
- Signal dated by source or discovery feed
This timestamp places the signal in the radar. It may be the discovery post date rather than the primary page's first publication date.
- Shared brief stored
AINewsIQ prepared one abstract from fetched source page.
A publication revision history is not stored for this signal. New engagement snapshots do not mean the source itself changed.
Why this may matter
If the reported storage and speed gains hold, internalizing memory in model computation could make long-history agentic workloads cheaper to run than retrieval or full KV caching, while preserving general capabilities.
Limits and open questions
Source claims not independently checked; the efficiency figures are the authors' self-reported results.
Organization-specific financial, security, and operational effects require the organization's own context and review.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the original source and published comments now.