01 · OBSERVE
Use the existing camera path
Keep the installed cameras, CVR recording, motion detection, object tags, and MotionGi video processing unchanged.
Proposed CVR edge-memory pilot
Turn motion clips into remembered action.
Cloudastructure already owns the edge-to-cloud video path. Engrammatic adds lightweight temporal memory on the CVR CPU, linking motion and object metadata into richer episodes before detailed cloud analysis.
6 weeks1 CVR8–25 camerasCPU firstshadow-mode safe
Cloudastructure × Engrammatic · edge-memory pilot2:31 · English narration + captions
The context gap
The CVR sees a stream. The cloud receives pieces.
A CVR can manage 5–50 cameras, perform motion and light person/vehicle detection locally, and forward motion clips for deeper cloud analysis. That saves bandwidth, but detailed analysis starts from separated fragments. The proposed sidecar keeps a cheap memory of the full metadata stream at the one place where it is still continuous.
02 · ENCODE
Read metadata, not pixels
Encode camera, time, motion, person or vehicle, zone, duration, and local track ID into one binary hypervector.
03 · REMEMBER
Link events across time
XOR and popcount compare each event with recent sequence memory, exposing unusual combinations and broken patterns.
04 · ESCALATE
Send one richer episode
Attach pre-roll, relevant motion clips, a novelty score, and reason codes for Cloudastructure’s cloud AI and operators.
Clear division of labor
Add temporal memory. Keep the stack authoritative.
Engrammatic is not a replacement for computer vision, the VMS, remote guarding, or Cloudastructure’s cloud analytics. It is a narrow, inexpensive decision layer between local detection and cloud escalation.
Cloudastructure CVR
video pathrecord + buffer
local signalmotion + objects
video efficiencyMotionGi
Engrammatic sidecar
inputevent metadata
operationXOR + popcount
outputepisode + reasons
Cloud AI + operators
visiondeep analysis
workflowverify + respond
authorityfinal decision
Product boundary: Engrammatic does not ingest raw pixels for this pilot and does not claim to identify people or vehicles. It remembers and scores the lightweight detections Cloudastructure already produces.
Device + cost answer
Software first. Hardware after profiling.
The pilot should answer the device question with measurements rather than a speculative appliance. Start on the existing CVR CPU. Preserve Jetson for heavier vision workloads. Add a separate box only if isolation or capacity data requires one.
Deploy the memory service as an add-on to the CVR CPU. It consumes compact event metadata and uses binary operations designed for commodity processors.
CPUexisting CVR capacityGPU / Jetsonleft available for visionRolloutcontainer or native service
If the CVR profile shows insufficient headroom—or operational isolation is preferred—the same software runs on a small x86 add-on. The BOM and unit cost are quoted only after CPU, RAM, thermal, and camera-density profiling.
Live pilot
Six weeks to a measurable go / no-go.
Run the new memory layer in shadow mode before it influences any upload. Compare today’s motion-clip path with memory-guided episodes against the same replay and live windows.
WEEK 1 · INTEGRATE
Agree the event contract
Map CVR metadata, camera topology, motion boundaries, local track IDs, clip retrieval, and one representative replay set.
WEEKS 2–3 · SHADOW
Score without control
Build baseline memories per site and camera group. Log every linked episode and reason code without changing uploads.
WEEKS 4–5 · COMPARE
Test two paths
Measure the current motion-clip path against memory-guided episodes at matched security-event recall.
WEEK 6 · DECIDE
Profile and recommend
Deliver the scorecard, failure analysis, CPU/RAM profile, unit-economics input, and product recommendation.
≥25%Target reduction in deep cloud-analysis calls
MATCHIncident recall versus the agreed baseline
<10%Target average CVR CPU overhead
ZEROCamera interruption attributable to the pilot
P95Edge scoring latency reported by camera density
BYTESUploaded volume per camera-hour
CONTEXTPre-roll + action + follow-on completeness
REASONSOperator-useful explanation coverage
Proposed pass gate: at least 25% fewer deep cloud-analysis calls at matched incident recall, under 10% average CVR CPU overhead, and zero camera interruption. These are validation targets to confirm in week one—not completed benchmark results.
Pilot outputs
Leave with evidence and a product decision.
The work is designed to answer the technical, device, commercial, and marketing questions together. The result is useful even if the pilot says not to proceed.
CVR event adapter + replay harness
A reproducible path from Cloudastructure’s lightweight detections into Engrammatic memory, with episode output and reason codes.
Live comparison report
Cloud calls, bandwidth, recall, context, latency, CPU, and RAM reported by camera density and event type.
Deployment and BOM recommendation
A measured decision between existing CVR CPU, an optional x86 sidecar, and the role Jetson should keep for vision.
Joint edge-memory story
This explainer and proposal page establish the concept. If targets hold, both teams approve a results case study and jointly marketed per-CVR SKU.
Cloudastructure · powered by Engrammatic memory
Give every CVR a memory of what happened between the clips.
Start with one recorder, one representative site, and an agreed replay set. Six weeks later, decide from measured recall, context, compute, and cloud workload.
Review the pilot scope →