# : Harnessing LLM Memory for Persistent RF Intelligence<\/p>\n\n\n\n
There is a fundamental limit to how much telemetry a human or an ephemeral system can process before it “over-tokens” its own context.<\/p>\n\n\n\n
In the world of Large Language Models (LLMs), we solve this with KV Cache compression. We identify which parts of the conversation are salient, which can be summarized, and which can be evicted without losing the “thread” of reasoning.<\/p>\n\n\n\n
In this development cycle, SCYTHE has weaponized these exact techniques for the cyber-physical domain.<\/p>\n\n\n\n
We are moving away from treating RF observations as disposable telemetry. We are starting to treat them as tokens in a behavioral context window.<\/p>\n\n\n\n
This is the beginning of Cognitive Cache Engineering for RF Scythe.<\/p>\n\n\n\n
—<\/p>\n\n\n\n
## The Problem: The Immortal Graph and the “Goldfish” Memory<\/p>\n\n\n\n
Until now, SCYTHE instances largely behaved like “field brains” with two modes of failure:<\/p>\n\n\n\n
1. Ephemerality: When the instance died, the longitudinal continuity of actors died with it.<\/p>\n\n\n\n
2. Over-saturation: If we kept everything, the graph became an unreadable hairball of “immortal” observations that lacked semantic hierarchy.<\/p>\n\n\n\n
Cognitive Cache Engineering solves both by implementing a Multi-Tier Semantic Memory hierarchy and Attention-Aware Pruning.<\/p>\n\n\n\n
—<\/p>\n\n\n\n
## \ud83c\udfdb\ufe0f Multi-Tier Semantic Memory: HOT, WARM, and COLD<\/p>\n\n\n\n
We have replaced our flat memory model with a tiered substrate that mirrors the biological memory consolidation process:<\/p>\n\n\n\n
– HOT (Active Clusters): Resides in the high-frequency `MacClusterEngine`. This is our “L1 Cache”\u2014active trajectories, live RF signatures, and current ASN bindings.<\/p>\n\n\n\n
– WARM (Recent History): When a cluster becomes inactive, it is demoted to WARM. It is summarized, but still available for sub-100ms Semantic Recall. If a “new” device appears that matches a WARM signature, it is instantly promoted back to HOT.<\/p>\n\n\n\n
– COLD (Archival): After an hour of inactivity, data is compressed into longitudinal actor memory. This isn’t just “storage”; it’s a retrieval-optimized archive of compressed behavioral primitives.<\/p>\n\n\n\n
—<\/p>\n\n\n\n
## \u2702\ufe0f Semantic Eviction: Attention-Aware Pruning<\/p>\n\n\n\n
We have abandoned simple time-based eviction (TTL). Age is a poor proxy for importance.<\/p>\n\n\n\n
Inspired by LLM Heavy Hitter Oracle (H2O) techniques, we now use a Retention Score to determine what stays in the cache. An entity’s “Attention Salience” is calculated based on:<\/p>\n\n\n\n
– Confidence: How well-grounded is this actor in the graph?<\/p>\n\n\n\n
– Novelty: Is this a new behavior or a known background signal?<\/p>\n\n\n\n
– Recurrence: Does this actor reappear across different temporal windows?<\/p>\n\n\n\n
– Threat Weight: Does the behavioral signature match known adversarial motifs?<\/p>\n\n\n\n
– Motion Consistency: Is the trajectory physically coherent?<\/p>\n\n\n\n
A stationary Starbucks AP decays in seconds. A rotating, locally-administered MAC moving between cellular towers remains “HOT” for hours, even if it goes silent.<\/p>\n\n\n\n
—<\/p>\n\n\n\n
## \ud83d\udcc9 Trajectory LoRA: Low-Rank Actor Compression<\/p>\n\n\n\n
Storing every GPS coordinate of a moving actor is inefficient and noisy. We have implemented Trajectory Compression\u2014essentially a “LoRA” (Low-Rank Adaptation) for physical motion.<\/p>\n\n\n\n
Instead of 4,000 raw observations, we store Motion Basis Vectors:<\/p>\n\n\n\n
– `stationary-periodic`<\/p>\n\n\n\n
– `linear-transit`<\/p>\n\n\n\n
– `vehicular-high-speed`<\/p>\n\n\n\n
This reduces the data footprint of actor history by 95%+ while actually *improving* prediction accuracy. Our DOMA (Dynamic Object Motion Analysis) model now reasons over these clean primitives rather than wading through raw coordinate jitter.<\/p>\n\n\n\n
—<\/p>\n\n\n\n
## \ud83d\udce1 Semantic Delta Streaming: Efficient Cognition<\/p>\n\n\n\n
The cost of visualization is often the bottleneck in tactical awareness. Sending full point-clouds or coordinate lists kills battery life and saturates links.<\/p>\n\n\n\n
Our new Semantic Delta Streaming protocol only sends the “Change in Meaning”:<\/p>\n\n\n\n
– Initial State: Send full basis and centroid.<\/p>\n\n\n\n
– Steady State: Send simple deltas (e.g., `{“delta”: [0.01, 0]}`).<\/p>\n\n\n\n
– Pivot: Send full state only when the motion basis changes or a spatial jump occurs.<\/p>\n\n\n\n
This results in a dramatic reduction in Cesium\/Deck.gl redraw pressure and extends the field-life of mobile units.<\/p>\n\n\n\n
—<\/p>\n\n\n\n
## \ud83c\udfd7\ufe0f The Cognition Hierarchy: Reflex to Strategy<\/p>\n\n\n\n
Finally, we have established an Embedding Hierarchy that maps dimensions to cognitive “depth”:<\/p>\n\n\n\n
1. Reflex (384-dim): Fast, cheap, always-on. Used for real-time MAC continuity and spatial stitching at the edge.<\/p>\n\n\n\n
2. Analytical (768-dim): Richer context. Used for actor attribution and identifying recurring campaign motifs.<\/p>\n\n\n\n
3. Strategic (LLM): Expensive and sparse. Used for generating the GraphOps Analyst narratives and high-level hypothesis generation.<\/p>\n\n\n\n
—<\/p>\n\n\n\n
## The Strategic Threshold: From Devices to Entities<\/p>\n\n\n\n
By applying LLM cache techniques to the RF domain, SCYTHE has crossed a critical threshold.<\/p>\n\n\n\n
We are no longer tracking devices (ephemeral MACs\/IPs). We are tracking behavioral entities.<\/p>\n\n\n\n
The graph is no longer a log; it is a continuously learning persistent world-model. Even if an adversary rotates every identifier they have, the “Cognitive Cache” remembers the latent identity hidden in the behavior.<\/p>\n\n\n\n
Phase 1 is complete. The foundation for a truly longitudinal, cyber-physical memory organism is now active.<\/p>\n\n\n\n