Liquid AI tunes shared agent memory for Snapdragon

Liquid AI has optimized its Liquid Context software for Qualcomm’s Hexagon neural processing unit. The collaboration targets phones, cars, wearables, and other Snapdragon devices, where the software can maintain a local user profile and supply relevant context to authorized agents running on the device, in the cloud, or across both environments.

Persistent personal context usually sits inside one vendor’s service or must be reconstructed during each session. Liquid proposes a device-level layer that multiple agents can access with the user’s permission. If device makers and agent developers adopt the interface, users could retain routines and preferences when they change apps, assistants, or hardware.

One profile, several agents

Liquid Context processes permitted device signals to infer routines, preferences, and current needs. It keeps that information in a shared context layer, then exposes selected details to connected agents instead of sending every raw signal to each service.

Liquid describes three examples of the proposed workflow:

• A message about a sick child prompts an agent to inspect the calendar, identify movable meetings, and draft rescheduling emails for approval.
• After a conference, an agent selects photos and drafts a LinkedIn recap using stored writing preferences.
• A smartwatch records a personal-best run and shares that context with a car, which adjusts the cabin temperature and suggests a nearby recovery stop.

These scenarios describe intended capabilities, not confirmed integrations in shipping devices. Cross-device examples would also require secure pairing, identity management, transport, and synchronization between the participating products.

Why Hexagon fits the workload

Maintaining context requires frequent processing as messages, location, activity, calendar events, and other signals change. Sending every update to a remote model would add network latency, cloud inference costs, and battery use while exposing more data outside the device.

Qualcomm’s Hexagon NPU combines scalar, vector, and matrix processing with hardware for transformer operations. Liquid says it has tuned its context memory layer and agent model for Snapdragon’s processing, memory, and latency characteristics, allowing much of the recurring work to remain on the device.

Local execution can reduce data transmission, but the architecture’s privacy properties will depend on access controls around the resulting profile. Per-agent permissions, encryption, retention limits, audit logs, deletion controls, and defenses against malicious context requests remain essential because an inferred profile can contain sensitive information even when the raw signals stay local.

Inside Liquid Agent

Liquid Agent uses LFM2.5-2.6B, a 2.6-billion-parameter model built for constrained hardware. Its hybrid architecture contains 16 blocks: 10 double-gated short-convolution blocks and six grouped-query-attention blocks. Short convolutions handle local patterns efficiently, while attention blocks capture relationships across longer spans of input.

Liquid reports that LFM2 delivers twice the CPU prefill and decoding performance of Qwen3. The company has not supplied enough benchmark detail in this announcement to compare hardware, model sizes, quantization levels, context lengths, or output quality. Liquid has separately reported 213 tokens per second for a 230-million-parameter sibling on a Samsung Galaxy S25 Ultra, though that smaller model is not the 2.6-billion-parameter model powering Liquid Agent.

Leverage on both sides

Liquid AI is an MIT CSAIL spinout founded by CEO Ramin Hasani, CTO Mathias Lechner, CSO Alexander Amini, and Daniela Rus. The company develops compact models for phones, vehicles, and other hardware with limited memory, power, and thermal capacity.

• Qualcomm gains a personalization layer that can demonstrate what Snapdragon’s NPU does beyond benchmark throughput.
• Liquid AI gains access to Qualcomm’s hardware expertise and a path to device makers building on Snapdragon.
• Device makers could add persistent personalization while retaining flexibility over which agents consume the context.
• Agent developers could receive structured user context without building a separate long-term profile for every device.

The collaboration alone does not place Liquid Context on every Snapdragon product. Qualcomm supplies silicon and software platforms, while manufacturers decide which components ship, what permissions they receive, and how deeply they integrate with the operating system.

A challenge to assistant lock-in

A device-level context service could weaken an assistant provider’s control over personalization. Apple, Google, Samsung, and other platform vendors currently tie much of their user context to their own accounts, services, and assistants. A neutral layer would let an authorized agent request the same underlying preferences and routines.

The design could also separate model choice from memory. A manufacturer might replace an agent model, route a task to a cloud service, or support several assistants while preserving the user’s context on the device. That flexibility depends on a stable context schema and permission model that outside developers can implement without privileged access.

What developers still need

Liquid and Qualcomm have announced the collaboration, and key implementation details for a generally available product remain unpublished:

• Availability: supported Snapdragon generations, device categories, launch dates, OEM partners, and geographic limits.
• Developer interfaces: SDKs, APIs, context schemas, retrieval methods, event subscriptions, quotas, and latency targets.
• Permissions: rules for granting, reviewing, revoking, and auditing each agent’s access to specific context.
• Data handling: encryption, retention periods, deletion behavior, backup policies, and cloud fallback conditions.
• Security: isolation between agents and protections against prompt injection, profile poisoning, impersonation, and excessive inference.
• Cross-device operation: pairing, authentication, synchronization, conflict resolution, and offline behavior.
• Performance: memory use, storage growth, NPU occupancy, battery impact, thermal behavior, and model quantization.
• Commercial terms: licensing, pricing, update responsibility, certification, and long-term support.

The adoption test

Liquid Context will require deep OEM integration to access device signals, run continuously, and serve multiple applications safely. It will also need third-party agent providers to trust an interface they do not control. Shipping devices, public developer tools, and documented security boundaries will show whether the collaboration becomes a reusable Snapdragon capability or remains limited to selected demonstrations and partner applications.