Dun & Bradstreet has spent over 180 years building a comprehensive commercial database. Its Commercial Graph, covering 642 million businesses and their relationships, corporate hierarchies and risk profiles, was designed for people. Credit analysts, risk managers and sales professionals who could wait for query results and work through ambiguous entity matches. AI agents cannot do any of those things.

When D&B’s customers started pushing agents into credit, procurement and supply chain workflows, the Commercial Graph that had reliably served nearly 200,000 customers globally became a problem. The systems built to serve human analysts were the wrong architecture for machines. So D&B rebuilt.

“We need to think about agents as our new consumer category, evolving from our standard credit analysts or sales and marketing professionals, et cetera, to also now catering to these customers’ agents,” Gary Kotovets, Chief Data and Analytics Officer at Dun & Bradstreet, told VentureBeat.

What broke when agents started querying

The Commercial Graph was not a single database. It was a collection of separate systems built for different use cases and different markets, held together by custom integrations. Human analysts navigated that fragmentation through SQL queries or pre-built interfaces. Agents could not.

The scale of the underlying data compounded the problem. The database had nearly doubled in five years, expanding from more than 300 million to more than 642 million business records, with 11,000 fields per record, according to D&B. The firm now runs approximately 100 billion data quality checks per month as records move through its systems. Querying that at the sub-second latency agents require, against a fragmented architecture, was not workable.

The relationships the graph tracked were also the wrong kind. Legacy systems recorded static connections between entities. A CEO was linked to a company. That was the line. Agents working on credit assessments or third-party risk need dynamic relationships: when that CEO leaves for a new company, which organization does their track record follow? When a subsidiary changes ownership, how does that propagate across a corporate hierarchy? Those questions required custom analyst work before. Agents cannot wait for custom analyst work.

The broader problem is not unique to D&B. Kotovets said he has spoken with hundreds of CDOs and CIOs over the past six months and consistently heard the same constraint: they could not build what they wanted in AI because their data foundations were not standardized, normalized or agent-queryable. D&B had that foundation, built over decades to serve human analysts. It still had to rebuild for agents.

What they actually built

The rebuild started with consolidation. D&B migrated its fragmented databases to cloud infrastructure, redesigned the underlying schema and built a data fabric layer that normalizes records across markets while preserving regional compliance requirements. The result is a unified knowledge graph that tracks billions of relationships across 642 million companies, continuously updated and enriched by AI-driven data processing.

On top of that graph, D&B built a structured access layer for agents. Raw SQL access at agent query volumes and latency requirements was not the answer. Instead, D&B created a set of tools and skills available through MCP that package data with context and route agents to the right records for specific queries. A match and entity resolution engine sits behind every query, confirming that when an agent asks about a company, the answer resolves to a verified, specific entity rather than a name match.

D&B solved agent identity from both directions

Rebuilding the graph and adding MCP access solved the data retrieval problem. It did not solve the identity problem. Agents are not humans, and the authentication model built for human users did not extend to machines.

D&B built a new registration model for agents. They must map to a verified IP address and register an individual access key, treated as an authenticated identity in the same pipeline as a human user.

“We actually have a concept of Know Your Agent, similar to know your customer, that does those additional verifications,” Kotovets said.

That handles the inbound problem: knowing which company an agent belongs to and what data it is entitled to query. But D&B also built for the outbound problem: what happens when a customer’s own multi-agent workflow loses track of which company it is analyzing.

In a workflow that chains a credit check agent, a KYC agent and a third-party risk agent, each queries D&B at a different step. Without a mechanism to confirm they are all referencing the same entity, a workflow can complete while operating on divergent records.

“They have to come back to our verification agent to ensure that they’re still talking to each other about the same entity,” Kotovets said. “It’s almost like a digital handshake, in a sense.”

D&B’s business verification agent can be embedded into any workflow as a persistent reference point and is available on Google’s A2A protocol regardless of which orchestration tool a customer uses.

Four things enterprises must get right before deploying AI agents

The rebuild exposed requirements that go beyond D&B’s own stack.

1. Data foundations come before agent infrastructure. The CDOs and CIOs Kotovets spoke with over the past six months consistently hit the same wall: they cannot build what they want in AI until their data is clean, normalized and consolidated. D&B had that foundation already. Most enterprises do not, and they will feel it.

2. Design for dynamic relationships, not static ones. Enterprise data systems typically record point-in-time connections: a person belongs to a company, an asset belongs to a subsidiary. Agents working on credit, risk or supply chain decisions need to reason across relationships that shift over time. If the underlying data only captures the static line, the agent will too.

3. Build entity consistency checks into multi-agent workflows. When multiple agents touch the same entity at different steps, there is no guarantee they are all referencing the same record by the time the workflow completes. That gap needs to be engineered for explicitly. Entity verification is a workflow design requirement, not an optional guardrail.

4. Embed lineage from the start, not as an afterthought. Every agent-produced answer should carry a traceable path back to its source. In credit, risk and supply chain decisions, the cost of an error is concrete. Lineage needs to be built in before scaling, not added after problems surface.

“You could always click and see where it came from, and validate it all the way back to the original source,” Kotovets said. “That’s been the key for us in unlocking a lot of other capabilities, because we have that level of certainty in the things that we’ve done.”

The London-based beauty and wellness booking platform has joined the UK unicorn club at a $1bn-plus valuation, in a deal that lands while the broader SaaS complex is busy arguing about its own funeral.

Fresha, the London-based booking and payments platform for salons and spas, has raised $80m from funds managed by KKR in a deal that values the company at more than $1bn, the company said on Thursday.

The round, structured as primary growth capital, takes Fresha to unicorn status and lifts the total raised since 2015 to $285m.

The cheque comes from KKR’s Next Generation Technology Growth fund, the firm’s growth-equity arm, which writes into companies that are already past the product-market-fit stage and are looking for scale capital rather than runway.

The numbers Fresha disclosed alongside the announcement explain the appetite. The platform is used by more than 130,000 beauty and wellness businesses across the UK, Australasia, the Gulf, North America and parts of South-East Asia, and processes more than 35 million appointments a month, or roughly 420 million a year, against $15bn in annual gross merchandise value.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

Annual revenue run-rate stands at more than $140m, growing at over 60% a year, and the business is profitable. The last time Fresha disclosed a valuation, in a Series C extension in late 2021, the figure was $640m.

Founded in 2015 by William Zeqiri and Nick Miller, Fresha has spent the past five years quietly displacing the older booking incumbents in its core markets and pushing into payments, capital, and, more recently, AI-driven scheduling and marketing tools.

Zeqiri, in a statement, called reaching unicorn status “a proud milestone” and said the round would fund further global expansion and AI investment.

Miller, the company’s chief product officer, framed the round as validation from customers, who he said were already using the platform as their primary operating layer.

KKR’s diligence ran for more than a year and included surveys of over 1,000 beauty and wellness businesses across the US, UK, Ireland, the EU and Australia, plus interviews with customers, former employees and competitors.

The research, according to KKR, ranked Fresha first across software quality, ease of use, support, set-up and marketplace strength, with an average score of 8.1 out of 10 against a competitor average of 6.7.

Patrick Devine, a partner on KKR’s Tech Growth team, said Fresha had built “a differentiated platform combining software, financial services and marketplace capabilities with embedded AI.”

Marta Szczerba, a director on the same team, said she had followed the founders for years and had been “highly impressed with the consistent performance.”

The deal lands at an awkward moment for the SaaS category Fresha sits inside. Salesforce is down roughly 30% year-to-date and the broader software complex has spent 2026 absorbing the argument that per-seat pricing is the wrong shape for the AI era.

A vertical platform earning revenue from payments and marketplace fees as well as subscriptions is, on the face of it, the kind of business that argument is least worried about, which is presumably part of what KKR’s diligence team concluded.

Fresha said the new capital will go toward expansion in the US, continental Europe, Africa and South-East Asia, and toward AI features across booking automation, marketing, accounting and workforce management. The company did not disclose a planned timeline to an IPO or any further fundraising.

Cyber-Crime

MFA? No problem, says crimeware that tricks users into handing attackers the keys to M365

The FBI has issued a public service announcement warning about a new phishing kit that’s stealing Microsoft OAuth tokens at an alarming rate.

OAuth token theft is a serious headache for organizations because stolen tokens can bypass multi-factor authentication (MFA) and grant access to privileged accounts within an organization without needing to know their credentials. 

Think corporate espionage, data theft, maybe even ransomware.

The main culprit is Kali365, described as a phishing-as-a-service platform that’s being peddled on Telegram, first spotted by crimefighters in April 2026.

“Kali365 lowers the barrier of entry, providing less-technical attackers access to AI-generated phishing lures, automated campaign templates, real-time targeted individual/entity tracking dashboards, and OAuth token capture capabilities,” the FBI said in its announcement.

Phishing kits aren’t new. Different flavors are always in development, but the good ones can be especially problematic for organizations.

Kali365 lets attackers send convincing phishing emails that impersonate “trusted cloud productivity and document-sharing services,” – Adobe Acrobat Sign, DocuSign, and SharePoint – according to security shop Arctic Wolf.

That email contains a device code and instructions for the target to enter the code into a legitimate Microsoft page, a hyperlink for which is included in the email.

Entering that code registers the attacker’s device to the unwitting target’s M365 account, effectively surrendering access to emails, Teams, and all the rest of it. No MFA required.

Arctic Wolf published a deep dive on Kali365 back in April, noting that it also offers adversary-in-the-middle (AitM) capabilities that are distinct from the device code phishing described by the FBI.

The second attack Kali365 enables leads to the same outcome, accessing Microsoft accounts while bypassing MFA, just through slightly different mechanics.

Victims are sent an initial phishing email containing a cookie-based lure, which transparently proxies their browser via attacker-controlled infrastructure, Arctic Wolf said. Requests are then forwarded to a real Microsoft login page, and responses are beamed back to the victim, who authenticates the typical way using their valid credentials, passing Microsoft MFA.

Session cookies, related artifacts, and other session information are scooped up during this process and stored in the Kali365 attacker panel. From there, attackers can generate scripts to replay those sessions in their own environment, effectively borrowing the genuine user’s session.

The researchers’ analysis of Kali365 revealed three distinct tiers for subscribers.

The lowest Client Tier is for individual attackers, who can change the branding on the panels to give each a bespoke look while sporting the same underlying powers. The Agent Tier is for resellers who can provision and manage their own branded Kali365 panels and Client Tiers. The Admin Tier is reserved for Kali365’s developers.

Kali365 has a simple pricing structure: $250 per month per tenant, or $2,000 for a year. It supports an array of languages: Arabic, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, and Turkish.

Since emerging in April, Kali365 has often been mentioned in the same breath as EvilTokens, another device code phishing platform that hit headlines weeks earlier after Microsoft confirmed hundreds of compromises each day.

“Each campaign is distributed at scale, targeting hundreds of organizations with highly varied and unique payloads, making pattern-based detection more challenging,” Tanmay Ganacharya, VP of security research at Microsoft, told The Register.

“We continue to observe high-volume activity, with hundreds of compromises occurring daily across affected environments.”

Both Arctic Wolf and the FBI suggested organizations at risk should use conditional access policies to block device code flow where not required.

Defenders should also consider blocking authentication transfer policies, which let users move authentication between devices such as PCs and phones.  ®

Samsung could finally be preparing the compact flagship Galaxy fans have been asking for.

A new leak suggests the company is working on a Galaxy S27 Pro model that would bring some of the Ultra’s premium features into a noticeably smaller phone.

According to reports from Korea, Samsung is developing a new 6.47-inch Galaxy S27 variant that would sit between the larger Galaxy S27 Plus and Ultra models alongside the standard Galaxy S27. If accurate, it would mark the first time Samsung has used this specific screen size in its flagship lineup.

More importantly, though, the Galaxy S27 Pro reportedly won’t just be another mid-tier option with a fancier name. The leak claims it could share “most” of the Galaxy S27 Ultra’s specs. However, it may drop features like the S Pen to keep the device smaller and more manageable.

That would make it a pretty interesting shift for Samsung. Right now, if you want the company’s best cameras, battery tech, and performance upgrades, you’re usually stuck buying the biggest phone in the lineup. The rumoured S27 Pro sounds closer to Google’s recent Pixel strategy. The Pixel 10 Pro offers nearly all the same flagship features as the larger Pro XL. However, it is in a more pocket-friendly size.

The reported 6.47-inch display would still make the S27 Pro larger than the regular Galaxy S26, which sits at 6.3 inches. However, it would be noticeably smaller than the Plus and Ultra models. That could hit a sweet spot for users who want flagship specs without carrying around something that feels tablet-adjacent.

Of course, there will probably be some compromises. A smaller phone usually means less internal space. Unless Samsung adopts newer silicon-carbon battery technology by then, battery capacity could end up being one of the trade-offs compared to the Ultra.

Still, this leak feels more believable than some of the earlier “Pro” rumours surrounding the Galaxy S27 series, largely because the market has shifted. Compact premium phones are quietly making a comeback, and Samsung doesn’t really have a true answer to devices like the Pixel 10 Pro yet.

There’s no official confirmation from Samsung for now. However, if the Galaxy S27 Pro does happen, it could end up being the most interesting model in the lineup — not the Ultra.

Looking for a new flagship smartphone but are torn between iOS and Android? You’ve come to the right place.

With both the iPhone 17 and Pixel 10 Pro sporting plenty of AI features, flagship processors and brilliant cameras, choosing between the two can feel like a challenge. Fortunately, we’ve reviewed both and have compared our experiences below.

Keep reading to see how the iPhone 17 compares to the Pixel 10 Pro. If you’re not sold on either, make sure you visit our best smartphones guide, where we’ve listed our favourite iOS and Android models across all budgets.

Alternatively, our Google Pixel 10 vs Pixel 10 Pro compares key models from Google’s flagship line-up.

Price and Availability

Both the iPhone 17 and Pixel 10 Pro are readily available to buy now. The iPhone 17 has a cheaper starting RRP of £799/$799 and comes in a choice of five colours: White, Black, Lavender, Sage and Mist Blue.

SQUIRREL_PLAYLIST_10207955

In comparison, the Pixel 10 Pro starts at £999/$999 although it can be found with solid price cuts. For example, at the time of writing you can pick up the Pixel 10 Pro for just £749 on Amazon. Otherwise, the handset comes in a choice of four colours: Moonstone, Jade, Obsidian and Porcelain.

SQUIRREL_PLAYLIST_10207881

Design

• Both the iPhone 17 and Pixel 10 Pro look similar to their respective predecessor – and that’s not necessarily a bad thing
• iPhone 17 is fitted with Action and Camera Control buttons
• Pixel 10 Pro supports Pixelsnap

The iPhone 17 looks similar to last year’s iPhone 16, and many other entry-level iPhones that came before it. However, this really isn’t a bad thing as the iPhone 17 is a sleek and well-designed handset, with flat edges and rounded corners that many of the best smartphones now sport. Plus, the five colour options give the iPhone that extra bit of personality too. 

Much like its predecessor, the iPhone 17 is fitted with Action and Camera Control buttons. The Action button sits just above the volume rockers and can be customised to act as a shortcut to quickly open apps, while Camera Control works as a shortcut to the camera. 

Similarly, the Pixel 10 Pro looks remarkably similar to the Pixel 9 Pro, and sports the same pill-shaped camera bar that looks so much sleeker than the Pixel 8 Pro’s own. In fact, the biggest difference between the Pixel 10 Pro and Pixel 9 Pro is the inclusion of Pixelsnap which supports Qi2 wireless charging.

The fact Google kept things similar this year isn’t a bad thing at all, as the handset feels great in hand and looks brilliant too.

Otherwise, both the iPhone 17 and Pixel 10 Pro are equipped with an IP68 rating which protects the handsets from dust and submersion in water.

Winner: Both are well-built and look great, but we’ll give the win to iPhone 17

Screen

• Both have 6.3-inch displays
• Pixel 10 Pro can get a bit brighter at 3300 nits compared to the iPhone 17’s peak of 3000 nits
• Both screens are hard to fault

Ranking the iPhone 17 and Pixel 10 Pro’s respective displays is no easy task as both are packed with plenty of premium screen technologies, including an LTPO 1-120Hz refresh rate (the first for the entry-level iPhone), impressively high peak brightness levels and crisp resolutions too. However, we should note that the Pixel 10 Pro can reach a slightly higher peak brightness of 3300 nits while the iPhone 17’s peak is 3000 nits. Even so, the difference is negligible. 

Not only that, but both are fitted with screen protection too which promises to offer scratch and drop resistance too. While the iPhone 17 sports Apple’s own Ceramic Shield 2, the Pixel 10 Pro is covered by Gorilla Glass Victus 2 instead. 

Essentially, we’ve concluded that the iPhone 17 boasts the “best screen yet on an entry-level iPhone” while the Pixel 10 Pro has the title of owning “one of the best phone screens around”. 

Winner: Tie

Camera

• Pixel 10 Pro has a dedicated 48MP telephoto lens
• iPhone 17 has an 18MP square selfie camera, which is a huge upgrade from last year
• Pixel 10 Pro is fitted with Google’s AI Camera Coach which offers photography advice

The iPhone 17 is fitted with just two rear lenses, including a 48MP main and 48MP ultrawide. While its main lens does have a 2x in-sensor zoom that delivers good quality shots, and can even be pushed to around 4x before detail is lost, if you want a dedicated telephoto lens then you’ll be better off with the Pixel 10 Pro or the iPhone 17 Pro instead.

Having said that, we think the iPhone 17’s dual camera is enough to suit most people. Its main lens delivers sharp, colour-accurate images, even in low-light conditions. While its companion ultrawide lens isn’t quite as reliable in tougher conditions, it also manages to capture the likes of scenic vistas and even macro shots with ease too.

However, the key upgrade with the iPhone 17’s overall camera set-up is arguably with the new 18MP selfie lens. While the jump from 12MP to 18MP might not sound too exciting, the key difference is the new square sensor that allows you to shoot portrait and landscape shots without needing to rotate the phone.

Otherwise, Google’s phones have earned the title of being among the best camera phones, and fortunately the Pixel 10 Pro continues this trend – albeit with some slight issues that should be kept in mind. 

With a 50MP main, 48MP ultrawide and 48MP 5x telephoto camera set-up, the Pixel 10 Pro reliably takes a good photo in most lighting conditions. Images look warm and rich during the day, while detail is preserved in night shots. While it does benefit from a dedicated telephoto lens, unlike the iPhone 17, zoom isn’t actually as reliable as we’d like. Push it past the 5x mark and shots feel artificial and too smooth, which suggests Google’s AI processing is at play. 

So, although it lacks the telephoto lens, we’d argue the iPhone 17 is more of a reliable camera phone overall.

Winner: iPhone 17

Performance

• Apple’s A19 vs Google’s Tensor G5 chips
• Both perform well in everyday use, but the iPhone 17 sees better benchmark results
• Pixel 10 Pro is built for AI capability rather than sheer power

We should start by saying that if you’re planning on playing demanding AAA titles or editing multiple 4K videos then neither the iPhone 17 nor the Pixel 10 Pro will suit you. Instead, you’re better off looking at more powerful options – perhaps one of the best gaming phones.

Otherwise, powering the iPhone 17 is Apple’s own A19 Bionic chip, while the Pixel 10 Pro runs on Google’s Tensor G5 processor instead. The iPhone 17 especially performs brilliantly in everyday use, with apps opening instantly and more casual titles playing admirably too. Plus, Apple has equipped the iPhone 17 with 256GB storage as a minimum, which is double the starting storage as the Pixel 10 Pro.

Similarly, the Pixel 10 Pro rarely sees any slowdown in day-to-day use, with apps and casual games opening and running quickly, although we did note that the camera has a slight lag to it, especially when shooting at 50MP. However, Tensor G5 is built with AI in mind so tends to prioritise performance in that area instead.

Winner: iPhone 17

AI features

• Google AI is easily one of the best AI toolkits, with genuinely useful features
• Apple Intelligence still feels like an afterthought

Following on from the above, the Pixel 10 Pro is well-equipped with plenty of Google’s AI features, which makes Apple Intelligence seem more like an afterthought in comparison. While some may go unused, tools like Circle to Search, Call Screening and Gemini are genuinely useful and come in handy. We especially found ourselves using Magic Cue more than we expected to, as it works behind the scenes but pops up when you’re likely to need it, say if you need directions to a restaurant you’ve booked.

We also find Google’s photo editing tools to be among the best, with the eraser reliably removing unwanted objects from the background. In comparison, Apple’s competing Clean Up feature often leaves you with pretty obvious signs that the photo has been tampered with.

In addition, other Apple Intelligence tools like Image Playground and Siri just don’t feel as smooth or thought through as Google’s own. With this in mind, if you want a phone that is designed to support AI, then the Pixel 10 Pro is a much easier recommendation.

Winner: Google Pixel 10 Pro

Software

• iPhone 17 runs on iOS 26 which is polished and easy-to-use
• Pixel 10 Pro sports Material 3 Expressive which is one of the nicest phone operating systems to use
• Google promises the Pixel 10 Pro will see seven years of updates but Apple doesn’t disclose the exact amount of support the iPhone 17 will see

The iPhone 17 runs on iOS 26 which saw the introduction of Liquid Glass, Apple’s new transparent UI design. While some weren’t keen on the redesign, we think it looks great and adds to the polished overall feel of iOS. 

We especially like how well iOS integrates with other Apple devices, and how easy it is to share files between the ecosystem. It’s a level of integration that Androids just can’t quite seem to match. 

This isn’t to say the Pixel 10 Pro’s software isn’t great. With Material 3 Expressive, the Pixel 10 Pro is easy-to-use, feels polished and offers plenty of customisation options that iOS doesn’t. Either way, both the iPhone 17 and Pixel 10 Pro boast brilliant software that we think you’ll struggle to fault.

Lastly, Google promises that the Pixel 10 Pro will see up to seven years of software updates, which will take you up to Android 23. While Apple doesn’t publicly disclose how many years of software updates the iPhone 17 will see, looking at previous years reveals it also offers around the seven year mark.

Winner: Pixel 10 Pro (as Google publicly announces its software updates)

Battery

• Neither are two-day handsets
• iPhone 17 supports faster wired charging at 40W compared to the Pixel 10 Pro’s 30W
• Both support Qi2 wireless charging

While many of the best Android phones can comfortably see two-days of battery life, unfortunately the Pixel 10 Pro doesn’t boast this claim. Instead, we found the Pixel 10 Pro can comfortably last one day with three or four hours of screen time while more demanding days with up to six hours would deplete the battery. 

Fortunately, with support for 30W wired and 25W (Qi2) wireless charging, topping up is that bit more convenient than the Pixel 9 Pro.

In comparison, we found the iPhone 17 could end a day with five hours of screen time with around 20% left in the tank. Finally, though it shares the same 25W (Qi2) wireless speeds, it bests the Pixel 10 Pro with 40W wired speed support instead.

Winner: iPhone 17

Verdict

Of course, deciding between the iPhone 17 and Google Pixel 10 Pro may simply boil down to your existing ecosystem. While it is perfectly possible to use an iPhone with a Windows PC, or a Pixel with a Mac, naturally using either handset with their own ecosystem offers much more of a seamless experience.

However, if you’re open minded then there’s a lot of factors to consider. The iPhone 17 boasts a more reliable camera set-up, despite its lack of telephoto lens, sees higher benchmark scores with its A19 processor and is cheaper than the Pixel 10 Pro. However, the Pixel 10 Pro’s AI toolkit is genuinely useful, and could improve how you use your phone on a day-to-day basis.

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.”

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.