MemT-1.0: Towards Memory-Native Evolution: plain AI paper summary
AI-generated summary. Read the original paper before relying on it.
Hugging Face paper page · Original paper on arXiv
MemT-1.0 puts long-term memory inside the model’s computation instead of carrying the full history as context. Its token-level updates and mixture of storage aim to retain procedural experience while keeping inference practical.
Key findings
- The design uses gated DeltaNet-2 and a mixture-of-Storage mechanism for updating memory.
- Training uses 34B tokens across the MemT-1.0 model family.
- The paper reports nano and Flash model variants.
- At 128K tokens, MemT-1.0-Nano reports about 1.86x faster end-to-end inference and 66x less persistent storage than a full-context key-value cache.
How you might use it
- Compare MemT with retrieval on repeated client procedures, checking whether it recalls the right instruction without carrying unrelated old work.
- Measure speed and storage while testing how an operator corrects an outdated memory.
Limits
The abstract does not specify model sizes, deployment hardware, or how memory errors are corrected or deleted.
Related research
- Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory This paper on arXiv Related paper on arXiv: Both papers address memory for persistent agents. MemT internalizes memory in model computation, while Mem0 provides a memory system for retaining useful information across interactions.
Business ideas
1. native memory pilot
Offer: A memory layer that records durable user facts with update provenance and correction history.
Buyer: Agent platform teams
Technical link: Use MemT token-level gated updates and mixture-of-storage as proposed controls.
Evidence
- competitor: LangSmith integrates with pytest, Vitest, and GitHub workflows for PR or nightly evaluations. (scope: vendor documentation year: 2026 region: global) (source)
- adoption: The proportion of UK businesses with 10 or more employees reporting use of at least one AI technology rose from around 12% to around 35%. (scope: ONS Business Insights and Conditions Survey analysis year: 2026 region: United Kingdom denominator: businesses with 10 or more employees) (source)
Differentiation: Arize Phoenix is the named memory-adjacent competitor; add retention evidence.
First test: Replay 100 support sessions with profile updates.
Success measure: Attribution remains correct on 90% of held-out questions.
Constraints: Pilot only. Use customer-owned or licensed data, approved runtimes, and human review. Market evidence is contextual proxy evidence rather than TAM.
2. memory update monitor
Offer: A monitor that shows which gated memory updates were accepted, rejected, or superseded.
Buyer: Agent platform teams
Technical link: Use MemT token-level gated updates and mixture-of-storage as proposed controls.
Evidence
- competitor: Arize Phoenix is open-source and local-first for tracing, evaluation, experimentation, and prompt iteration. (scope: vendor product description year: 2026 region: global) (source)
- adoption: The proportion of UK businesses with 10 or more employees reporting use of at least one AI technology rose from around 12% to around 35%. (scope: ONS Business Insights and Conditions Survey analysis year: 2026 region: United Kingdom denominator: businesses with 10 or more employees) (source)
Differentiation: Arize Phoenix is the named memory-adjacent competitor; add retention evidence.
First test: Apply 50 contradictory preference updates.
Success measure: The monitor flags 95% of conflicting updates.
Constraints: Pilot only. Use customer-owned or licensed data, approved runtimes, and human review. Market evidence is contextual proxy evidence rather than TAM.
3. storage tier planner
Offer: A storage planner that assigns memories to tiers based on recency and retrieval value.
Buyer: Agent platform teams
Technical link: Use MemT token-level gated updates and mixture-of-storage as proposed controls.
Evidence
- competitor: LangSmith integrates with pytest, Vitest, and GitHub workflows for PR or nightly evaluations. (scope: vendor documentation year: 2026 region: global) (source)
- adoption: The proportion of UK businesses with 10 or more employees reporting use of at least one AI technology rose from around 12% to around 35%. (scope: ONS Business Insights and Conditions Survey analysis year: 2026 region: United Kingdom denominator: businesses with 10 or more employees) (source)
Differentiation: Arize Phoenix is the named memory-adjacent competitor; add retention evidence.
First test: Load three storage tiers with 10,000 memory items.
Success measure: Retrieval latency stays under the agreed 200 ms budget.
Constraints: Pilot only. Use customer-owned or licensed data, approved runtimes, and human review. Market evidence is contextual proxy evidence rather than TAM.
4. longitudinal profile service
Offer: A profile service that answers longitudinal questions with citations to retained memories.
Buyer: Agent platform teams
Technical link: Use MemT token-level gated updates and mixture-of-storage as proposed controls.
Evidence
- competitor: Arize Phoenix is open-source and local-first for tracing, evaluation, experimentation, and prompt iteration. (scope: vendor product description year: 2026 region: global) (source)
- adoption: The proportion of UK businesses with 10 or more employees reporting use of at least one AI technology rose from around 12% to around 35%. (scope: ONS Business Insights and Conditions Survey analysis year: 2026 region: United Kingdom denominator: businesses with 10 or more employees) (source)
Differentiation: Arize Phoenix is the named memory-adjacent competitor; add retention evidence.
First test: Run 30 users through a month-long profile simulation.
Success measure: Relevant memories are available in 90% of sampled sessions.
Constraints: Pilot only. Use customer-owned or licensed data, approved runtimes, and human review. Market evidence is contextual proxy evidence rather than TAM.
5. memory retention audit
Offer: An audit that finds stale, duplicated, or unsupported memories before agent reuse.
Buyer: Agent platform teams
Technical link: Use MemT token-level gated updates and mixture-of-storage as proposed controls.
Evidence
- competitor: LangSmith integrates with pytest, Vitest, and GitHub workflows for PR or nightly evaluations. (scope: vendor documentation year: 2026 region: global) (source)
- adoption: The proportion of UK businesses with 10 or more employees reporting use of at least one AI technology rose from around 12% to around 35%. (scope: ONS Business Insights and Conditions Survey analysis year: 2026 region: United Kingdom denominator: businesses with 10 or more employees) (source)
Differentiation: Arize Phoenix is the named memory-adjacent competitor; add retention evidence.
First test: Audit 200 memories after expiry and correction events.
Success measure: The audit finds 95% of stale or unsupported memories before reuse.
Constraints: Pilot only. Use customer-owned or licensed data, approved runtimes, and human review. Market evidence is contextual proxy evidence rather than TAM.