AI

One day last October, sitting in the courtyard of his house in China’s Henan province, Dong Hui decided to see if he could hold a pen to write.  Dong, 39, had sustained spinal cord injuries in a car accident six years earlier that left him paralyzed from the neck down. Slowly but determinedly, he wrote his name, “Thank you,” and then the date. This was the result of an 11-month-long rehabilitation enabled by an implant in his brain. Before that process, Dong could move his arms slightly but wasn’t able to use his fingers. “I couldn’t believe I was able to write again. I was so excited I even missed a stroke in my name,” he told MIT Technology Review on a video call.  In November 2024, Dong became one of the first people in China to be given an invasive brain-computer interface (BCI) through brain surgery. He had signed up for a clinical trial with the device’s developer one month after seeing on TV how a BCI had apparently enabled another paralyzed Chinese man to hold his granddaughter.  This March, the implant Dong uses became the first invasive BCI product in the world to be approved for use beyond clinical trials. It’s now available to some patients with paralysis in their limbs due to spinal cord injuries. We spoke to a range of experts to understand why the device was able to reach this global milestone, what makes this moment so significant, and what to expect next.  A world first Dong’s brain implant is a coin-size device called NEO. It was developed by Neuracle Technology, a Shanghai-based startup, together with researchers at Tsinghua University in Beijing.  During a procedure that took just over an hour and a half, the device’s sensors, which collect Dong’s brain signals, were placed on his dura mater, the tough outer layer of tissue that covers and protects the brain. The signals are transmitted to a computer by an implant placed on Dong’s skull. The computer then translates the signals into commands for a soft robotic glove Dong wears during the 2.5-hour training sessions he completes each day to help him learn to grab.  Dong started his rehabilitation around a week after surgery. “On the ninth day of my training, my right hand successfully grabbed a ball without the glove,” he says. “That was a miraculous moment.”  Now he continues with his training at home. He wants to be able to control his hands better in order to put on clothes, eat, and do other daily tasks without troubling his aging parents.  A growing number of people with traumatic injuries in China are now poised to tread a similar path thanks to NEO’s recent approval. According to China’s National Medical Products Administration, the bureau responsible for drug supervision, the product is suitable for patients between 18 and 60 who have paralysis in all limbs due to spinal cord injuries but still have some residual function in their arms.  NEO beat several other BCIs to approval, including one from Neuralink, a California-based company founded by Elon Musk. Since October 2023, Neuracle has conducted 36 clinical trials using NEO, including the one on Dong. Thirty-two of them took place in the space of a few months in 2025, with the details about one of the four first in-person trials published in a preprint paper last July. Neuracle did not reply to a request for comment from MIT Technology Review. One reason for NEO’s fast approval could be that it has a “relatively less invasive” design than counterparts such as Neuralink’s N1 brain chip, says Avinash Singh, a BCI researcher at the University of Technology Sydney. NEO’s eight sensors sit on top of the brain’s protective membrane while Neuralink’s N1 chip directly penetrates the cortex, the outermost layer of the brain itself. Neuracle’s device faces fewer regulatory constraints because it presents a lower risk of hemorrhage, glial scarring, and long-term signal degradation, Singh says. China’s strong support for its BCI industry also means that NEO was put on an expedited regulatory pathway; in comparison, the approval process of the US Food and Drug Administration can take several years, Singh adds. A big boost for BCIs NEO’s approval is hugely important for the global BCI industry, says Wang Shouyan, a neuroscientist at Fudan University in Shanghai who was not involved in research or trialing for NEO. Even though research and development on BCIs has taken place for several decades, most of it happened in the lab. The news means that BCIs are now ready for large-scale manufacturing and clinical use in China, Wang says.  For Dong, however, it means something much more personal. “Now, it will be able to help not only me, but also thousands and thousands of other patients suffering from spinal cord injuries in China who are tortured by despair each day,” he says of NEO. “It will bring them hope and change their lives.”  Days after NEO was approved, China started incorporating it into the country’s health insurance system by assigning it a unique code. This is one of the first steps toward a future where eligible Chinese patients pay a certain percentage of the BCI’s price if they need it during their treatment. The growth of China’s BCI industry is expected to accelerate thanks to the government’s policy support and financial backing. The country’s latest five-year plan, published on the same day Neuracle received its approval, lists BCI as one of six key industries important to China’s future tech competitiveness, alongside quantum technology, humanoid robots, and others. Several Chinese startups, including NeuroXess and StairMed, have already worked in the field for many years.  “China’s decision to double down on becoming a global leader in the field owes in part to what these companies have already accomplished,” says Meicen Sun, an information scientist at the University of Illinois Urbana-Champaign who studies information and technology policy.  But, Sun says, the biggest advantage China may have is that Chinese people, particularly patients like Dong, tend to welcome

Hermes Agent already remembers across sessions. The open-source agent from Nous Research ships with curated memory files and full-text session search. But a new community project argues that built-in memory is too shallow for serious work. A new library named ‘Memory OS‘ has been released under an MIT license by a developer (ClaudioDrews). It stacks six memory layers onto Hermes. It adds a vector database, structured facts, and an auto-curated knowledge wiki. The project is new but it seems to have a good potential and its architecture shows how agent memory can be layered. Memory OS Memory OS is not a Hermes plugin you toggle on. It is a layered system that sits beside Hermes Agent’s own memory. Hermes already provides workspace files and a session database. Memory OS keeps those and adds four more layers above them. The full stack runs locally using Docker, Qdrant, Redis, and Python 3.11+. It works with any LLM provider Hermes supports, including OpenRouter, OpenAI, Anthropic, and Ollama. The README frames it as a “memory operating system,” not a single feature. The Six Layers, From Files to Vectors Layer 1 is Workspace. It holds MEMORY.md, USER.md, and CREATIVE.md, injected into the system prompt each turn. Layer 2 is Sessions. It uses state.db, a SQLite database with FTS5 full-text search across conversation history. Layer 3 is Structured Facts. It stores durable facts in memory_store.db, using SQLite, HRR, FTS5, and trust scoring. A feedback loop adjusts those trust scores over time, alongside entity resolution. Layer 4 is Fabric, a heavily forked version of the Icarus Plugin. This fork adds LLM-powered session extraction over the upstream esaradev/icarus-plugin. It handles cross-session recall through 16 tools, including fabric_recall, fabric_write, and fabric_brief. Layer 5 is the Vector Database, built on Qdrant. It uses 4096d Cosine vectors plus BM25 sparse search, a keyword-style ranking method. Layer 6 is an LLM Wiki, an auto-curated vault of concepts, entities, and comparisons. That wiki is continuously ingested back into Qdrant through a process called wiki-continuous-ingest. How the Retrieval Flow Works The flow sits on when memory is read and written. On pre_llm_call, Memory OS runs what it calls surgical recall. It pulls from four sources at once: Fabric, Qdrant, Sessions, and Facts. Each source is gated by a relevance threshold before anything reaches the model. Per-session deduplication stops the same context from appearing twice. A social-closer filter skips trivial messages, such as a plain “thanks.” On post_llm_call and on_session_end, the system extracts and captures new learnings automatically. The stated goal is token efficiency, not stuffing the context window. The Fallback Cascade and Cleanup Layer 5’s retrieval uses a four-level fallback. It tries hybrid search first, then dense vectors, then lexical, then SQLite. If one method fails or returns nothing, the next takes over. This design keeps recall working even when the vector database struggles. Memory OS also runs a weekly decay scanner to age out stale entries. Semantic dedup merges near-identical memories when cosine similarity exceeds 0.92. These housekeeping steps aim to stop memory from bloating over months of use. Local-First, And Deliberately So Memory OS positions itself against cloud memory services like mem0, Zep, and Letta. Its pitch is that memory infrastructure should run on your own machine. The memory data stays local, with no memory subscription. LLM calls still go to whichever provider you choose. Hermes itself already supports eight external memory providers, including mem0 and Honcho. Memory OS is not one of those official providers. It is a separate, community-built stack layered on Hermes directly. For teams with data-residency rules, a local memory store can matter. Just open-sourced **Memory OS** — a complete hierarchical persistent memory architecture for the Hermes Agent. 6 layers, fully local:• Structured facts + trust scoring with feedback loop• Hybrid vector search (Qdrant + BM25)• Self-curating LLM Wiki• Semantic… — Claudio Drews (@ClaudioDrews25) May 31, 2026 Strengths and Limitations Strengths: Clear layered design separating files, sessions, facts, vectors, and a wiki Fully local infrastructure with no cloud memory subscription Provider-agnostic, matching Hermes Agent’s own flexibility Token-efficient retrieval by design, via gated sources and per-session deduplication Limitations: Brand new, with few commits A forked Icarus Plugin that the author says is not upstream-compatible Heavier setup: Docker, Qdrant, Redis, and an ARQ Worker all required No published benchmarks on recall quality, latency, or token savings Key Takeaways Memory OS is a community-built, MIT-licensed stack that adds six memory layers on top of Hermes Agent. It combines workspace files, FTS5 session search, trust-scored facts, a forked Icarus fabric, Qdrant vectors, and an auto-curated LLM wiki. Retrieval runs on pre_llm_call with gated, deduplicated recall from four sources; capture runs on post_llm_call and on_session_end. Memory infrastructure is fully local and provider-agnostic, but LLM calls still go to your chosen provider. Check out the Repo. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Meet Memory OS: A 6-Layer Open-Source Memory Stack Built on Top of Hermes Agent appeared first on MarkTechPost.

Trajectory’s concurrent multi-LoRA stack reports a 2.81× experiment-throughput gain over single-tenant RL, with all code in the NovaSky-AI/SkyRL GitHub repository. Most language models improve in discontinuous jumps. A team collects data, trains, and ships a new version. This takes months and produces remarkable or catastrophic behavior for users. Trajectory wants to replace that cycle with continual learning. The Trajectory team published a field report describing how. It built a concurrent, multi-LoRA training platform for continuously learning workloads. The work was done with UC Berkeley Sky Lab and Anyscale. All training code is open-sourced in the NovaSky-AI/SkyRL repository. The result is a 2.81× end-to-end experiment-throughput improvement. The comparison is against a single-tenant training framework. Trajectory reports no regression on any training rewards. What Multi-LoRA Training Actually Is Continual learning requires models to update from live feedback and production interactions. A coding agent could learn engineering patterns as developers correct its work. A support agent could resolve hard tickets as operators intervene on difficult cases. Most training infrastructure still assumes a linear lifecycle. Teams allocate GPUs, initialize the model, run a job, then spin down. Continual learning revises that relationship. When production interactions become training inputs, training becomes part of a live system. Modern RL training reduces to three core primitives. The Sampler generates trajectories from the current policy model. The Trainer computes gradients and updates the policy weights. Parameter synchronization broadcasts updated weights back to inference workers. Trajectory calls its approach Continuous Multi-LoRA Training, or C-LoRA. Each experiment maps to a dedicated LoRA adapter on a warm, multi-tenant engine. The Problems It Targets The Trajectory team identifies four inefficiencies in traditional stacks: (1) Cold starts are slow: Every serial job reloads checkpoints, initializes the distributed runtime, and warms inference engines. For large models, this step alone can exceed 30 minutes per run. (2) RL is memory intensive: Frontier models often exceed 100B parameters. Qwen3.5-397B can require up to eight H200 nodes to fit into memory. LoRA cuts memory usage by an order of magnitude. It freezes the base model and trains only small adapter weights. (3) Traditional stacks are single-tenant: They run one experiment at a time. Multi-LoRA maps each experiment to one adapter, multiplexing throughput by a factor of N. (4) Job utilization is low: Trainers and inference engines stall while waiting for each other. Multi-LoRA load balances across jobs to fill idle capacity. Inside the Architecture Most throughput wins come from inference. In vLLM, all adapters are hot-loaded in GPU memory. Decode steps can then mix tokens from different adapters in the same batch. The key enabler is the SGMV decode kernel. It fuses per-adapter matrix-vector work into one GPU launch per decode step. After each optimization step, updated LoRA weights load in-place into the inference engine. The scheduler does not freeze, so other tenants keep decoding. Training works differently. One active LoRA adapter trains on the GPU. The rest sit in pinned CPU memory. Each tenant’s state lives in an AdapterStore. It holds LoRA parameters, FP32 master weights, optimizer moments, and gradient buffers. The engine swaps one tenant’s state onto the GPU, runs a single forward_backward pass, then swaps it back. This training path is still single-adapter. The inference concurrency gains do not yet apply to training. The Numbers Trajectory tested on a single H200 node with Qwen3-4B-Instruct-2507. It ran sync RL on GSM8K in an agentic setting. The Trajectory team reframed GSM8K as a tool use learning task. The model decides when to call a Calculator and a Final Answer tool. Reward is 1.0 only when Final Answer is called with the correct answer. The policy starts near 40% accuracy at step 0. With the right learning algorithm, it climbs past 90% by step 9. The Trajectory team scaled to eight concurrent multi-LoRA runs. Final Experiment Time hit 5433s at N=8, a 2.81× speedup. Eight concurrent experiments finished before three serial runs back-to-back. Mean Experiment Time also improved, peaking at N=4 with a 1.88× speedup. Every concurrency level reached reward_accuracy above 90% by step 9. The Tradeoffs Higher throughput costs per-step latency. As N grows, First Experiment Time and Step Time degrade. At N=8, the first serial experiment finishes 1.97× faster. Mean step time rises from 191s to 500s, only 2.62× slower. Most of that increase is rollout time. Rollout grows from 162s to 401s, roughly 77% of the increase. At N=2, doubling the load adds only 15% rollout time. That is the ideal case for multi-LoRA. The pattern held on a harder workload. On τ-bench retail with the NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 MoE model, N=2 finished 10 steps 1.28× faster. Per-tenant step time rose 1.57×. Strengths and Weaknesses Strengths: 2.81× end-to-end experiment-throughput gain at eight concurrent runs No accuracy regression; runs tracked the serial baseline within ±1σ in the final steps LoRA cuts memory by an order of magnitude versus full fine-tuning Fully open-sourced in NovaSky-AI/SkyRL for the community to build on Weaknesses: Per-step latency and First Experiment Time degrade as N grows Training remains serialized across tenants; only inference is multiplexed Tested mainly on mid-sized models, not frontier-scale parameters Setup requires an 8× H100/H200 node and a Megatron build Key Takeaways Trajectory built a concurrent, multi-LoRA RL training stack for continual learning, open-sourced in NovaSky-AI/SkyRL. It reports a 2.81× end-to-end experiment-throughput gain over a single-tenant baseline, with no reward regression. Each experiment maps to a dedicated LoRA adapter on an always-hot engine, multiplexing throughput by N. Most gains come from vLLM multi-LoRA inference via the SGMV decode kernel; training stays single-adapter. The tradeoff is per-step latency: at N=8, step time rises from 191s to 500s. Marktechpost’s Visual Explainer Field Report · May 27, 2026 Continuous Multi-LoRA Training for Continual Learning Trajectory, built with UC Berkeley Sky Lab and Anyscale. 2.81× end-to-end experiment-throughput gain Training code open-sourced in the NovaSky-AI/SkyRL repository. 01 — What it is One always-hot engine, many adapters Continual learning updates models from live feedback and production interactions. Trajectory calls its approach Continuous Multi-LoRA Training (C-LoRA). Each experiment maps to a dedicated LoRA adapter on

Text-to-speech TTS moved fast over the past year. The line between synthetic and human speech narrowed. Latency dropped below 100 milliseconds for some real-time systems. Emotional control became a standard feature rather than a research demo. This guide reviews the models that really matter in 2026. It is written for AI professionals choosing a model for production. How to read TTS benchmarks in 2026 Two benchmarks dominate in most community discussions. The first is the Artificial Analysis Speech Arena Leaderboard. It ranks models by blind human preference using an ELO rating. As of 2026 it evaluates dozens of production APIs. The second is the community-run TTS Arena on Hugging Face. It uses the same blind A/B voting method. These leaderboards measure perceived quality, not accuracy. They also change continuously. As of May 30, 2026, the Artificial Analysis Speech Arena lists Gemini 3.1 Flash TTS, Realtime TTS-2 (Research Preview), Sonic 3.5, Realtime TTS 1.5 Max, and Fun-Realtime-TTS-Preview as its top five by ELO. Those positions shifted within the prior weeks, and they will shift again. Treat any single number as a point-in-time reading, not a fixed truth. Accuracy needs separate measurement. Trelis Research tested ten models using a round-trip character error rate, or CER. The method transcribes generated audio with an ASR model, then compares it to the input text. Mean opinion score, or MOS, captures perceived naturalness. Both metrics have limits. Round-trip CER depends on the ASR model’s own accuracy. The UTMOS quality estimator was trained on audio up to ten seconds, so longer samples show less score spread. Latency is the third axis. The relevant figure for voice agents is time-to-first-audio, or TTFA. Time-to-first-byte, or TTFB, can be misleading, since container headers carry no audio. Consistency matters as much as the median. A Gradium benchmark from May 2026 measured the interquartile range across providers. Tail latency, not the average, determines user experience at scale. In short, no benchmark is complete. Quality, accuracy, latency, language coverage, and price all trade off. The right model depends on which axis your application cannot compromise. Commercial leaders #1 Inworld TTS-1.5 and Realtime TTS-2 Inworld AI is a research lab founded by a team from Google and DeepMind. It released TTS-1.5 on January 21, 2026. The model targets real-time, consumer-scale applications. Inworld reports roughly 30 percent more expressive range than TTS-1. It also reports about 40 percent better stability, measured through word error rate and output consistency. TTS-1.5 ships in two tiers. The Mini tier is tuned for latency-sensitive workloads such as voice agents and gaming. The Max tier balances higher stability with low latency. Inworld reports P90 time-to-first-audio under 130 milliseconds for Mini and under 250 milliseconds for Max. The model supports 15 languages and offers both instant and professional voice cloning. Pricing is tiered by plan, not a single rate. On the On-Demand and Creator plans, Inworld lists $25 per million characters for TTS 1.5 Mini and $35 for Realtime TTS-2 and TTS 1.5 Max. The Developer and Growth plans cut those rates; Growth reaches $15 for Mini and $25 for Max and TTS-2. Enterprise pricing goes as low as $5 and $10 respectively. Note that TTS 1.5 covers 15 languages, while TTS-2 covers over 100. Inworld later added Realtime TTS-2 in 2026. It is described as a closed-loop voice model with stronger steering and expressiveness. Across several leaderboard snapshots, Inworld reported holding three of the top five spots on the Artificial Analysis Speech Arena. Inworld suits developers building voice agents at consumer scale. The combination of low latency and aggressive pricing is its main draw. #2 Google Gemini 3.1 Flash TTS Google DeepMind released Gemini 3.1 Flash TTS on April 15, 2026. It is a preview model available through the Gemini API, Google AI Studio, Vertex AI, and Google Vids. The model introduces more than 200 audio tags. These tags steer style, tone, pacing, accent, and scene direction. On Google’s own report, the model reached an ELO of 1,211 on the Artificial Analysis leaderboard. It supports 70-plus languages and native multi-speaker dialogue. Google built it on the Gemini family rather than a standalone speech stack. The model treats generation as a language task: it decides not only what to say, but how to say it. The model has documented limitations that matter for deployment. A TTS session has a 32,000-token context window, and Google’s docs state that Gemini TTS does not support streaming. It is built for controlled text recitation, not interactive voice agents; the separate Live API is Google’s real-time path. Output quality can drift on generations longer than a few minutes, so Google recommends chunking. The model offers 30 prebuilt voices. All generated audio carries a SynthID watermark for AI-content identification. Gemini 3.1 Flash TTS fits podcast and audiobook generation with fine-grained control. It is a strong default for teams already on Google Cloud. #3 ElevenLabs v3 ElevenLabs released Eleven v3 in alpha on June 5, 2025. It reached general availability in early 2026, per the company’s announcement. ElevenLabs describes it as its most expressive model. It introduced inline audio tags formatted in lowercase square brackets. Examples include [whispers], [laughs], [sighs], and scene cues like [interrupting]. The model supports more than 70 languages. The GA release refined the alpha. ElevenLabs reports users preferred the new version about 72 percent of the time. It also improved how the model handles numbers, symbols, and specialized notation. A key feature is Text to Dialogue. It weaves multiple voices into one generation pass. The model matches prosody and emotional range across speakers. It can handle interruptions and shifting moods with limited prompting. Eleven v3 still requires more prompt engineering than earlier models. It is not built for real-time use. ElevenLabs states the larger model and higher-fidelity codec take longer to run. For real-time and conversational use, the company recommends Flash v2.5 instead. Those models stream with low latency, around the 75-millisecond range in vendor figures. ElevenLabs v3 fits narrative content, audiobooks, and character work where quality outweighs speed. It remains a common