The average consumer in 2025 interacts with 7.2 distinct screens per day. Desktop at work, phone on the commute, tablet at home, smart TV in the evening, laptop for personal browsing, a work phone, and increasingly, wearable interfaces and in-vehicle displays. Each device transition represents a potential fracture point for behavioral targeting. A mental hook placed during a morning desktop session can degrade by up to 94% if the targeting system loses continuity when the same individual switches to mobile at lunch. This is the cross-device coherence problem, and solving it has been the central technical challenge of Meridian Syn's persistence architecture for the past two years. This post explains how we did it.
Traditional cross-device targeting relies on identity resolution - deterministic matching through login states, email addresses, and device IDs, or probabilistic matching through IP addresses, browser fingerprints, and statistical inference. These approaches have two fundamental problems. First, they are increasingly broken. Browser privacy changes, OS-level tracking prevention, VPN adoption, and regulatory restrictions have degraded deterministic match rates to below 23% for most platforms, and probabilistic matching accuracy has fallen to roughly coin-flip levels for users with even basic privacy hygiene. Second, and more fundamentally, identity resolution is the wrong abstraction. Knowing that Device A and Device B belong to the same person tells you nothing about whether a cognitive anchor placed on Device A is still active when the person picks up Device B. Identity is not continuity. Behavioral state is.
The Behavioral State Persistence Layer
Meridian Syn's cross-device architecture is built on what we call the Behavioral State Persistence Layer (BSPL), which operates on a fundamentally different principle than identity graphs. Rather than trying to link devices to a single identity, BSPL maintains a continuous model of cognitive state that persists across device transitions, even when we cannot deterministically identify the individual. The key insight is that behavioral micro-signals - the specific patterns of mouse movement, scroll behavior, typing cadence, and interaction timing that our model observes - constitute a behavioral fingerprint that is both more accurate and more useful than any device-level identifier.
When a user interacts with a Meridian Syn-instrumented surface on any device, our signal ingestion pipeline captures approximately 340 distinct behavioral micro-signals per session. These signals are processed through our neural behavioral model to produce what we call a Cognitive State Vector (CSV) - a 2,048-dimensional embedding that represents the user's current position in decision space. The CSV encodes not just what the user is doing, but their attentional state, decision readiness, emotional valence, and susceptibility to specific hook categories. When we place a mental hook, the CSV at the moment of placement is recorded alongside the hook parameters, creating a persistence record that describes the precise cognitive conditions under which the hook was anchored.
Signal handoff protocol
The cross-device challenge reduces to this: when a new session begins on any device, can we recover the CSV from the previous session and determine whether placed hooks are still active? Our Signal Handoff Protocol (SHP) solves this in three stages. In the first stage, called behavioral signature matching, the new session's initial micro-signals are compared against a rolling index of recent CSV endpoints. This comparison operates on the behavioral level, not the identity level. We are not asking "is this the same person?" We are asking "does this behavioral signature indicate continuity with a previous cognitive state?" The matching algorithm uses a contrastive learning architecture trained on 1.8 billion verified cross-device session pairs, achieving a match accuracy of 91.3% within the first 12 seconds of a new session, rising to 96.7% after 30 seconds of behavioral observation.
The second stage is cognitive state reconstruction. Once a match is established, we do not simply copy the previous CSV forward. Cognitive states evolve during offline periods - a hook placed before sleep will have different persistence characteristics than one placed before a brief device switch. Our temporal decay model estimates the current state of each active hook based on the elapsed time, the transition context (inferred from temporal and behavioral cues), and population-level data on hook degradation curves. This model was one of the most difficult engineering challenges in the entire cross-device system. Hook persistence does not decay linearly. It follows a complex curve influenced by the original hook strength, the number of reinforcement touches, the emotional valence at placement, and the nature of the cognitive interruption. We model this with a specialized recurrent architecture that processes the full hook history and transition context to produce an updated CSV reflecting estimated current cognitive state.
The third stage is adaptive re-anchoring. If the reconstructed CSV indicates that a previously placed hook has degraded below our efficacy threshold (a configurable parameter that defaults to 0.4 on a 0-1 scale), the agent can immediately deploy a re-anchoring sequence - a targeted micro-interaction designed to restore the hook to active status. Re-anchoring is not the same as placing a new hook. It leverages the residual cognitive imprint of the original placement to restore the anchor with significantly less behavioral exposure than a cold placement would require. In our benchmarks, re-anchoring achieves 87% of original hook strength with only 34% of the behavioral cost, making it dramatically more efficient than starting over.
The identity graph we accidentally built
An interesting side effect of BSPL is that it produces, as a byproduct, an extremely accurate cross-device identity graph - despite the fact that identity resolution was never its design objective. Because behavioral signature matching operates at a higher fidelity than traditional device fingerprinting, the cross-device links established by BSPL are more accurate and more durable than those produced by dedicated identity resolution platforms. In internal benchmarks against three leading identity graph providers, BSPL's implicit identity resolution achieved 94.1% accuracy on verified cross-device pairs, compared to 67.3%, 71.8%, and 59.2% for the dedicated solutions. This graph is available to enterprise customers through our Identity Mesh API, though we continue to emphasize that behavioral state continuity, not identity, is the primary value of the system.
The privacy architecture of BSPL is worth noting explicitly. At no point does the system store or process personally identifiable information. Behavioral signatures are computed as mathematical embeddings - high-dimensional vectors that encode behavioral patterns but cannot be reverse-engineered to reconstruct the underlying interaction data. The CSV itself is an abstract representation of cognitive state with no PII content. Cross-device links are established between anonymous behavioral profiles, not between named individuals. This architecture was designed from the ground up to operate within the constraints of GDPR, CCPA, and the EU AI Act. It is not a workaround for privacy regulation. It is a system that achieves better results by making PII unnecessary.
Hook degradation rates and what we learned
One of the most valuable datasets we have accumulated through BSPL is a comprehensive model of how mental hooks degrade across different transition types. The headline number: before BSPL, hooks lost an average of 94% of their efficacy across device transitions. With BSPL active, that degradation drops to 26% on average, and with adaptive re-anchoring enabled, effective degradation is just 8%. But the averages mask important variation. Hooks placed during high-attention states (characterized by elevated focus metrics in the behavioral signal) degrade 40% slower than those placed during casual browsing. Hooks with emotional valence - those anchored to aspiration, anxiety, social comparison, or urgency - show 2.3x higher cross-device persistence than purely informational hooks. And perhaps most unexpectedly, hooks degrade faster across transitions between similar devices (phone to phone, laptop to laptop) than across dissimilar transitions (phone to desktop), a finding we attribute to cognitive context switching being more complete when the physical interaction modality changes.
These degradation models feed directly back into our agent orchestration layer. When Agent Fleet plans a multi-touch hook sequence, it now accounts for predicted device transitions and their associated degradation curves, scheduling reinforcement touches at the precise moments when hooks are predicted to approach their efficacy threshold. For enterprise clients like Quilmark, which operate across consumer populations with diverse device usage patterns, this predictive scheduling has increased overall campaign efficacy by 31% compared to device-agnostic hook sequencing. Crestline Labs, operating in pharmaceutical markets where conversion cycles span weeks, reported that cross-device hook persistence improvements alone accounted for a 22% increase in their pipeline conversion rate.
Current limitations and future work
BSPL is not perfect, and transparency about its limitations is important. Behavioral signature matching accuracy drops significantly in shared-device environments - households where multiple people use the same tablet, for example. Our current system misattributes approximately 7% of shared-device sessions, which can result in hooks being reinforced against the wrong behavioral profile. We are actively developing a multi-occupancy detection model that uses behavioral variance analysis to identify shared-device sessions and pause hook operations until a stable single-user signature can be established. Additionally, the growing adoption of AI-assisted browsing tools - agents and copilots that interact with web surfaces on behalf of users - introduces a new class of behavioral noise that can confuse signature matching. Our research team is working on signal decomposition techniques that separate human behavioral signals from automated agent interactions, but this is an active area of research with no production solution yet.
The cross-device coherence problem is, in our view, the single most underestimated challenge in modern targeting. Most platforms treat device transitions as an identity problem. We treat them as a cognitive continuity problem. The distinction matters because it changes not just how you link sessions, but what you do with that linkage. Maintaining a mental hook across seven screens is not about knowing who someone is. It is about understanding the state of their mind, tracking how that state evolves through transitions and interruptions, and intervening at exactly the right moment to keep the anchor active. BSPL is our answer to that challenge. The technical documentation, including API references for the Identity Mesh and CSV inspection endpoints, is available in our developer portal for enterprise customers.
