AI agents forget. Every time a coding assistant loses track of a debugging thread, or a data analysis agent re-ingests the same context it already processed, the team pays in latency, token costs, and brittle workflows. The fix most teams reach for — expanding the context window or adding more RAG — is increasingly expensive and still doesn’t reliably work.

To address this, researchers from Mind Lab and several universities proposed delta-mem, an efficient technique that compresses the model’s historical information into a dynamically updated matrix without changing the model itself. The resulting module adds just 0.12% of the backbone model’s parameters — compared to 76.40% for one leading alternative — while outperforming it on memory-heavy benchmarks. Delta-mem allows models to continuously accumulate and reuse historical data, reducing the reliance on massive context windows or complex external retrieval modules for behavioral continuity.

The long memory challenge

The conventional solution is to simply dump all the information into the model’s context window.

But as Jingdi Lei, co-author of the paper, told VentureBeat, current systems treat memory merely as a context-management problem. “Either we keep expanding the context window, or we retrieve more documents through RAG,” Lei explained. “These approaches are useful and will remain important, but they become increasingly expensive and brittle when agents need to operate over long-running, multi-step interactions, and they don’t really [work] like human memory since they are more like looking up documents.”

In enterprise settings, the bottleneck is not just whether the model can access history, but whether it can reuse that history efficiently, continuously, and with low latency. Standard attention mechanisms incur a quadratic computational cost as the sequence length increases. Furthermore, expanding the context window does not guarantee the model will actually recall the information effectively. Models often suffer from context degradation or context rot as they become overwhelmed with more (and often conflicting) information, even if they support one million tokens in theory.

The researchers argue for advanced memory mechanisms that can represent historical information compactly and maintain it dynamically across interactions. Existing solutions come with heavy trade-offs and generally fall into three paradigms:

• Textual memory: stores history as text injected into context — constrained by window limits and prone to information loss under compression.

• Outside-channel (RAG): encodes and retrieves from external modules — adds latency, integration complexity, and potential misalignment with the backbone.

• Parametric: encodes memory into model weights via adapters — static after training, can’t adapt to new information during live interactions.

Inside delta-mem

To achieve a compact and dynamically updated memory, delta-mem compresses an agent’s past interactions into an “online state of associative memory” (OSAM). This state is maintained as a fixed-size matrix that preserves historical information while the underlying language model remains frozen.

For enterprise workflows, this translates directly to resolving operational bottlenecks. Lei noted that a persistent coding assistant, for example, “may need to remember project conventions, recent debugging steps, user preferences, or intermediate decisions across a workflow.” Similarly, a data analysis agent might “need to maintain task state, assumptions, and prior observations while iterating over multiple tool calls.” 

Rather than repeatedly retrieving and re-inserting all relevant history for these tasks, the delta-mem matrix provides a low-overhead way to carry forward useful interaction states inside the model’s forward computation.

During generation, the system does not retrieve raw text segments to add to the prompt. Instead, the backbone LLM’s current hidden state is projected into the matrix to retrieve old memory. This operation extracts context-relevant associative memory signals from delta-mem. These signals are then transformed into numerical corrections that are applied to the computations of the model. This steers the model’s reasoning at inference time without altering its internal parameters.

Following each interaction, delta-mem updates the online state using “delta-rule learning.” When new information arrives, the previous state makes a prediction about the resulting attention values. It then compares this prediction to the actual value and corrects the memory matrix based on the discrepancy.

This update mechanism relies on a “gated delta-rule.” Basically, the memory module has different knobs that control how much previous memory is kept and how much of the new memory is applied. This error correction with controlled forgetting allows the matrix to evolve over time, holding onto stable historical associations without being derailed by short-term noise.

The researchers explored three strategies for determining when and how the matrix updates:

• Token-state write captures fine-grained changes but is vulnerable to short-term noise.

• Sequence-state write averages tokens within a message segment, smoothing updates at the cost of some localized detail.

• Multi-state write decomposes memory into sub-states for different information types like facts or task progress.

Delta-mem in action

The researchers evaluated delta-mem across three LLM backbones: Qwen3-8B, Qwen3-4B-Instruct, and SmolLM3-3B. They configured the framework with a compact 8×8 matrix. The system was tested on general capability benchmarks, including HotpotQA, GPQA-Diamond, and IFEval. It was also evaluated on memory-heavy tasks such as LoCoMo, which tests long-term conversational memory, and Memory Agent Bench, which assesses retention, retrieval, selective forgetting, and test-time learning over extended interactions.

The framework was compared against representative models from the three existing memory paradigms: textual memory baselines (e.g., BM25 RAG, LLMLingua-2, and MemoryBank), parametric systems (Context2LoRA and MemGen), and the outside-channel approach MLP Memory.

Across the board, delta-mem outperformed the baselines, according to the researchers. On the Qwen3-4B-Instruct backbone, the token-state write variant achieved an average score of 51.66%, easily surpassing the frozen vanilla backbone at 46.79% and the strongest baseline, Context2LoRA, at 44.90%. On the memory-heavy Memory Agent Bench, the average score jumped from 29.54% to 38.85%. Performance on the specific test-time learning subtask nearly doubled from 26.14 to 50.50.

However, the most compelling takeaways are the system’s operational efficiency. The researchers tested the framework in a no-context setting where the historical text was entirely removed from the context. Even without explicit text replay, delta-mem successfully recovered context-relevant evidence in multi-hop tasks. The researchers argue that the model remembers past interactions without needing to ingest massive amounts of prompt tokens.

The framework also adds only 4.87 million trainable parameters, representing just 0.12% of the Qwen3-4B-Instruct backbone. By comparison, the MLP Memory baseline required 3 billion parameters, scaling up to 76.40% of the backbone’s size while delivering inferior results. When prompt lengths scaled up to 32,000 tokens during inference tests, the framework maintained almost the exact same GPU memory footprint as a standard, unmodified model. It sidesteps the heavy memory bloat that affects other advanced memory systems like MemGen and MLP Memory.

Different update strategies proved beneficial depending on the underlying model capacity. The sequence-state write strategy was the most effective for stronger backbones like Qwen3-8B. These more capable models use the segment-level writing to smooth out updates and mitigate token-level noise. Conversely, the multi-state write strategy drove massive performance leaps for smaller backbones like SmolLM3-3B. For these lower-capacity models, separating memory into multiple states proved critical to minimizing information interference.

Implementing delta-mem in the enterprise stack

The researchers have released the code for delta-mem on GitHub and the weights for their trained adapters on Hugging Face. For AI engineering teams looking to integrate this framework into their existing inference stack, the process requires minimal computing resources.

“In practice, an engineering team would start from an existing instruction-tuned backbone, attach the Delta-Mem adapter modules to selected attention layers, train only the adapter parameters on domain-relevant multi-turn or long-context data… and then run inference with the memory state updated online during interaction,” Lei said. Crucially, teams do not need a massive pretraining corpus. The training data only needs to reflect the target memory behavior, such as multi-turn dialogues, agent traces, or domain workflows where earlier information must influence later decisions.

While compressing interaction history into a fixed-size mathematical matrix creates immense efficiency, it does come with trade-offs. Delta-mem is not a lossless replacement for explicit text logs or document retrieval. Because different pieces of information compete inside the same limited state, there is a risk of memory blending.

“Delta-Mem is useful when the system needs fast, online, continuously updated behavioral state,” Lei said. “RAG is better when the system needs exact factual recall, citation, compliance, auditability, or access to a large external knowledge base.” Remembering a user’s working style or a multi-step reasoning trajectory is a perfect fit for delta-mem, while retrieving a legal contract or a medical guideline should remain in a vector database.

This means the most realistic enterprise architecture moving forward is a hybrid approach. Delta-mem acts as a lightweight internal working memory, reducing the need to retrieve or replay everything all the time, while RAG serves as the explicit, high-capacity memory layer.

“Looking ahead, I do not think vector databases will become obsolete,” Lei said. “Instead, I expect enterprise AI stacks to become more layered. We will likely see short-term working memory inside the model, longer-term explicit memory in retrieval systems, and policy or audit layers that decide what should be stored, retrieved, forgotten, or exposed to the user.”

Apple has shown off the new Accessibility features coming in iOS 27, which did nothing to stem the torrent of rumors about what we’ll see in Apple Intelligence, but possibly did steal a little bit of thunder from Google’s peculiar mishmash of an I/O conference, on the AppleInsider Podcast.

It’s surely the only time of the year where Apple actually tells us something in advance about the next version of iOS. No guessing, no rumors, just straight information about the new or improved accessibility features.

Apple does so solely because World Accessibility Day is coming up, and not at all because Google is running its I/O developer conference at this time. Just as it’s entirely coincidental that Apple issued invitations to its own WWDC now as well.

We all use Google, but even if you’re not an Android fan, it used to be interesting to watch Google I/O. You’d always see some features that you wished Apple would adopt, for instance, but not this year.

This year Google I/O was full of sound and fury, signifying nothing, if you spell sound and fury as “AI.” Nothing this year was biting at Apple’s heels, and that’s downright peculiar since Apple is relying on Google Gemini for its forthcoming updates to Apple Intelligence.

BONUS: Subscribe via Patreon or Apple Podcasts to hear AppleInsider+, the extended edition. This time, speaking of developer conferences, we’ve got WWDC on the horizon and just like you, that means we have plenty we want to see launch there.

Links from the show:

More AppleInsider podcasts

Tune in to our Smart Home Insider podcast covering the latest news, products, apps, and everything HomeKit related. Subscribe in Apple Podcasts, Overcast, or just search for HomeKit Insider wherever you get your podcasts.

Podcast artwork from Basic Apple Guy. Download the free wallpaper pack here.

Those interested in sponsoring the show can reach out to us at: [email protected].

Subscribe to AppleInsider on:

Keep up with everything Apple in the weekly AppleInsider Podcast. Just say, “Hey, Siri,” to your HomePod mini and ask for these podcasts, and our latest HomeKit Insider episode too. If you want an ad-free main AppleInsider Podcast experience, you can support the AppleInsider podcast by subscribing for $5 per month through Apple’s Podcasts app, or via Patreon if you prefer any other podcast player.

Finnish phone maker HMD today launched its first smartphone, called the Vibe 2 5G, which comes preloaded with Indian AI company Sarvam’s chatbot Indus. Both companies had first announced the partnership during the India AI summit held in New Delhi in February.

The Indus app is powered by Sarvam’s locally trained 105-billion-parameter model — a measure of the AI’s scale and sophistication — and launched at the AI summit. The app supports 22 Indic languages and mid-sentence code-switching (the ability to fluidly mix languages mid-conversation, like switching between Hindi and English), which helps the assistant better understand the context of a query. Currently, the application doesn’t support offline usage, and it doesn’t have any integrated feature with the device to invoke the AI assistant through a shortcut.

The partnership is a potential testing ground for both companies to gauge the appetite for an India-focused chatbot.

“With this partnership, the first thing we want to do is get the Indus app to consumers,” said Ravi Kunwar, HMD’s CEO and Vice President for India and APAC, in an interview with TechCrunch. “Once they start using it, we will move to phase two to focus on driving more traction and stickiness. Right now, by pre-loading the app, we want to be more accessible to users,” he said.

The Vibe 2 5G is a mid-range Android phone with a 6,000mAh battery and a price tag of ₹10,999($114). Kunwar added the devices in the Vibe series of smartphones will also get the chatbot, and the company is also expected to launch a feature phone with Sarvam AI integration in the coming months.

That feature phone integration may ultimately prove more significant for both companies. HMD held a 4% share of India’s feature phone market in 2025, but its smartphone share was negligible — the company doesn’t even appear in the top 15, according to analyst firm IDC.

While it’s early days for Indus, the download numbers reflect that. Nearly three months after its launch, the app has been downloaded just over 293,000 times in India across platforms, according to Appfigures. By comparison, ChatGPT was downloaded 43.9 million times in the country.

It’s a big gap, but the strategy behind the HMD deal may matter more than the early numbers. Bundling a regional AI assistant with affordable hardware — particularly feature phones — is one of the more direct distribution plays available in a market as large and linguistically diverse as India, where English-language AI tools have limited reach. For investors and operators watching how AI adoption gets seeded in emerging markets, this partnership is worth tracking.

Sarvam has been one of India’s marquee AI startups. Beyond the Indus app launch, the company has focused on enterprise partnerships, especially for voice-based solutions. It is on track to become one of the most funded AI startups in the country, with reports suggesting a funding round of $300 million at a $1.5 billion valuation is in the works.

from the where’ve-you-been-bill? dept

Just a few days ago, I wrote a post about how Bill Cassidy had been primaried out of returning as a senator for Louisiana and how all of this bootlicking of the Trump administration obviously didn’t do the job he hoped it would do. As a result, he has been left as a lame duck senator with a legacy that will be primarily about his decision to belay his own moral stances generally and his heavy hand in RFK Jr. leading HHS under Trump 2.0.

The point of that post was two-fold. First, I wanted to highlight just how damning to his legacy the appointment of Kennedy to HHS has become to his legacy. Second, I wanted to highlight that this supposedly serious senator was perfectly willing to give up on his principles the moment he thought, incorrectly as it turns out, that it would be politically expedient to do so.

And if you need a bow to put on that second point, you can get it now that Cassidy has flipped his vote on the Senate bill to end America’s involvement in the war with Iran until the Trump administration gets authorization from Congress.

Sen. Bill Cassidy, R-La., who just lost his primary for renomination over the weekend after he faced opposition from Trump, voted “yes” to advance the measure, the first time he has done so after having repeatedly voted “no.”

“While I support the administration’s efforts to dismantle Iran’s nuclear program, the White House and Pentagon have left Congress in the dark on Operation Epic Fury,” Cassidy said in a statement. “In Louisiana, I’ve heard from people, including President Trump’s supporters, who are concerned about this war. Until the administration provides clarity, no congressional authorization or extension can be justified.”

Advertisement

It’s amazing how post-election-loss clarity can assist someone in rediscovering their own spinal cord. Now, you can read Cassidy’s comments about how Congress has been left in the dark and that he’s hearing from people worried that maybe this whole warlord routine by Trump isn’t so great and believe that Cassidy came to all of these epiphanies in the last couple of days… if you want. But I’m going to point at you and laugh in your face if you do.

Now that Cassidy has nothing to lose, he’s decided to do the right thing. That isn’t some feather in his cap. It’s a self-indictment of all of his actions leading all the way up to his primary loss. If Cassidy thought this vote was the right thing to do today, what made it the wrong thing to do a week ago? The answer is nothing.

Even if a vote is taken and the bill passes, it would still need to get through the Republican House and survive a presidential veto. There is little chance of either happening. But that isn’t the point.

The point is that Bill Cassidy could have been a patriot over the past year and a half since Trump’s reelection, but he chose not to until he didn’t have a Senate seat to defend. And that makes him a coward.

Filed Under: bill cassidy, donald trump, iran, principles, rfk jr., spine

The US is betting big on quantum, committing $2bn in federal funding to nine companies as it races to build a domestic quantum computing industry.

The US Department of Commerce announced yesterday it is proposing $bn in federal incentives under the CHIPS and Science Act, targeting two domestic quantum foundry companies and seven quantum computing firms. The funds are aimed at solving the most critical technology challenges in the race to develop utility-scale, fault-tolerant quantum computers.

The largest single allocation goes to IBM, which is using the $1bn CHIPS award to launch Anderon, a standalone company and what it says will be the first pure-play US quantum chip foundry. Headquartered in Albany, New York, Anderon will operate as a 300-millimetre quantum wafer foundry serving multiple quantum hardware vendors.

IBM will match the government’s contribution with $1bn of its own cash, along with significant intellectual property, assets, and its skilled workforce, with additional investors expected as Anderon grows.

GlobalFoundries is receiving $375m to launch Quantum Technology Solutions, a new quantum business built on its decade-long cryogenic CMOS, advanced packaging, and materials science investment. The Malta, New York-based foundry will manufacture the full quantum hardware stack, from quantum processor units to cryogenic readout and control chips, serving multiple qubit architectures including superconducting, trapped ion, photonic, topological, and silicon spin. The new business launches with existing customers that include Diraq, PsiQuantum, Quantinuum, Google, Microsoft, and Nvidia.

Among the seven quantum computing companies receiving funding is Quantinuum, which has signed a letter of intent for $100m to fabricate low-loss integrated photonics and specialised optical components tuned to trapped-ion critical wavelengths. The company plans to partner with GlobalFoundries for critical semiconductor components, and Monarch Quantum for integrated photonics.

“These strategic quantum technology investments will build on our domestic industry, creating thousands of high-paying American jobs while advancing American quantum capabilities,” said US Commerce Secretary Howard Lutnick.

The CHIPS R&D Office said it is taking a portfolio approach to strengthen US leadership across multiple quantum modalities simultaneously, while focusing each award on discrete technological problems of real consequence. The remaining $538m in funding is distributed among Atom Computing, Diraq, D-Wave, Infleqtion, PsiQuantum, and Rigetti Computing.

Quantum computing, the department said, has significant implications for national defence, advanced materials and biopharmaceutical discovery, financial modelling, and energy systems. The announcement comes as Quantinuum is also preparing for a Nasdaq IPO.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Gartner expects 60% of AI projects that lack AI-ready data to be abandoned through 2026. “AI-ready” sounds like a technical label. In practice, it describes something much more ordinary: whether a business can actually find, combine, and use the data it already owns.

That is a data fragmentation problem: the state where the information you need to plan a business decision is scattered across different systems, formats, teams and legal entities.