A Seoul startup building AI agents that claim to run without the memory bottleneck plaguing today's large language models has quietly raised approximately $2.27 million in seed funding. TraceWeave closed the round with Mirae Asset Capital leading and NAVER D2SF participating, betting that its Phase-TITAN architecture can eliminate the key-value cache that forces Transformer models to consume tens of gigabytes at long context windows.
The announcement positions the investment less as a vote on incremental improvements and more as a wager on a fundamentally different approach. "On-device AI is a core technology across diverse device environments beyond PCs, and it protects privacy," said Sanghwan Yang, head of NAVER D2SF, in the company's release. NAVER D2SF previously backed Nota AI, which focuses on model compression. TraceWeave represents a complementary bet on replacing the attention mechanism entirely rather than simply making it smaller.
A Different Memory Model
Phase-TITAN represents intelligence as evolving state rather than accumulated cache, according to the company. The pitch is technical but straightforward: a 1-billion-parameter Phase-TITAN model holds roughly 0.04 gigabytes of state at a one-million-token context, while a comparable Transformer's KV cache balloons to tens of gigabytes under the same load, per data cited by Dealroom. TraceWeave describes the design as a "deterministic dynamical system" that sidesteps attention and the linear memory growth it brings.
The startup demonstrated three proof-of-concept applications in July 2024. A 33-minute interactive world model session ran at 512×512 resolution with a fixed state size around 10.5 megabytes. A training-free dynamics monitor, deployed without pre-trained weights, detected a real fixed-wing engine power loss while running online, according to video documentation on the company's site. TraceWeave also reported that Phase-TITAN matched or exceeded a Matrix Profile baseline across four fixed-wing fault types in the ALFA dataset, a 47-flight Carnegie Mellon collection with 23 engine failures. The tests measured volume under the surface of the precision-recall curve after warm-up, with two to 23 training flights per fault type.
Whether those benchmarks hold up under independent scrutiny remains an open question. The startup has not published peer-reviewed results or released model weights publicly. Standard evaluations on suites like RULER for long-context retrieval or BFCL for function calling, at matched training compute, have yet to appear.
The Team and Early Traction
Park Jae-young, co-founder and CEO, previously worked in machine learning before launching TraceWeave with co-founders Kim Tae-hoon (CTO) and Park Geun-cheol (research and engineering). The three-person team reached the top 20 in NVIDIA's Inception Grand Challenge, according to the Seoul Economic Daily. Antler Korea listed the startup in its Cohort 7 portfolio.
The pre-money valuation sat at approximately ₩30 billion (around $22 million), according to Dealroom.
A Crowded Field
TraceWeave enters a field that has grown unexpectedly crowded over the past two years. Mamba, published in late 2023 and named an Outstanding Paper at COLM 2024, introduced selective state-space models that report constant-memory inference. Mamba-2 formalized structured state-space duality and claimed two-to-eightfold training speedups at small-to-medium scale, according to Papers with Code. Microsoft Research released RetNet, pitching chunk-recurrent retention as a "successor to Transformer," while RWKV markets a hybrid Transformer-RNN with recurrent state for long contexts.
Commercial uptake has followed, perhaps more than the founders expected. NVIDIA shipped Nemotron-H, a hybrid Mamba-Transformer mixture-of-experts model. IBM built Mamba variants into Granite 4. Mistral launched Codestral 7B with Mamba layers. AI21 fielded Jamba. Technology Innovation Institute released Falcon-Mamba 7B, an ACM TIST survey noted. A recent arXiv preprint titled "KV Cache Is the New Memory Wall" argued that bandwidth-bound inference at long context makes the cache the binding constraint, reinforcing vendor interest in alternatives.
Other startups are chasing similar territory. Pathway is building what it calls a "post-Transformer frontier model" with sparse activity and continuous learning, securing AWS backing in a startup profile published this year. Auxerta runs a "post-transformer pretraining stack" for experiments, according to research notes on its site.
Device Economics Drive Interest
The memory argument arrives as device makers ship AI-capable endpoints at scale, though not always with the RAM to match ambitions. Counterpoint Research reported that GenAI-capable smartphones will reach 45 percent of global shipments in the coming year and 52 percent the year after, with cumulative GenAI phone volume topping one billion units. An IDC forecast warned that a memory shortage could depress PC shipments substantially this year and delay recovery, raising the stakes for architectures that hold memory flat.
Google retrofitted Gemini Nano with frozen multi-token prediction to accelerate on-device inference under tight RAM budgets. Microsoft's Copilot+ PC program requires at least 40 TOPS of NPU performance to run local features including Recall and real-time translation. Qualcomm CEO Cristiano Amon told a Snapdragon Summit audience that agents would dominate the current generation of devices, projecting shipments "from hundreds of millions to probably billions in a very short term."
NVIDIA published guidance for building in-vehicle AI agents, noting that DRIVE AGX and Thor platforms run seven-to-13-billion-parameter models locally with sub-500-millisecond latency. ABI Research data cited in the post forecasts vehicles with agentic AI will rise from roughly five million currently to 70 million within a decade. Google and Valtech outlined a Nexus software-defined-vehicle stack integrating Android Automotive with a Gemini-based agent, running hybrid edge-cloud.
Grand View Research pegged the on-device AI market at $10.7 billion this year and $75.5 billion within the next decade, a 27.8-percent compound annual growth rate. Mordor Intelligence estimated on-device AI inference software at roughly $4.6 billion currently, reaching $17.4 billion over the next several years.
Next Steps
"Taking the KV cache out was never the hard part," Park said in the release. "Staying capable without it was... The best part is no part." The company declined to break out hiring targets or product roadmap milestones.
TraceWeave currently pitches Phase-TITAN language models for "persistent language intelligence without transformer layers or a token-growing KV cache," targeting on-device, long-context, and tool-use scenarios. The company's site describes use cases in mobile devices, robotics, and automotive systems where memory constraints bite hardest.
What remains unclear is whether the architecture can scale to the sizes and capabilities that have made Transformers dominant, or whether constant memory comes at the cost of representational power. The startup's proof points are intriguing but limited. Industry observers will be watching whether TraceWeave can move from demos to deployed systems, and whether those systems can match the performance benchmarks that Transformer-based models have set across language, reasoning, and generation tasks.
The funding gives the team runway to find out. For now, TraceWeave is another entry in the post-Transformer sweepstakes, one more bet that the attention mechanism's memory appetite has become unsustainable as AI moves from cloud to edge.
