MARS (Memory for Autonomous Real-Time Systems) is a GPU-resident memory store whose retrieval API exposes three orthogonal ranking primitives as first-class kernel-level controls — scope, time, and cross-modal connectivity — for real-time embodied AI: autonomous vehicles, humanoid robots, AR/VR headsets, and voice agents. Existing GPU vector search libraries (FAISS-GPU, cuVS CAGRA) expose a single contract: rank by cosine similarity over the entire indexed corpus. Embodied workloads typically know more than that at query time — an active track id, a dialogue session id, a current sub-task, or a recency window. Encoding this as host-side filtering over a global ANN sweep wastes GPU work and, without an ID filter, loses recall at scale on tiny multimodal clusters (a known property of unfiltered ANN). MARS consolidates scope, temporal decay, cross-modal BFS, and streaming append into one device-resident kernel API. Defensible summary of the headline result: episode-scoped retrieval as one kernel-level contract at a latency within ~1.3× of filtered FAISS (IDSelectorRange, single prototype at N=1M) and 33× below the unfiltered exhaustive sweep that stacks without an ID filter run today. On the kids-ball benchmark the scoped path is near-flat at ~200 µs p99 from N=10K to N=1M. Three ranking primitives, one GPU kernel pipeline Episode-scoped retrieval (RetrievalScope::EpisodeScoped) — restricts Stage 1 to the CSR member range of a named episode and skips BFS. Near-flat ~200 µs p99 from N=10K to N=1M. With 10-node episodes and K=15, scoped hit@15 = 1.0 is guaranteed by construction (the scope is the answer set); the comparison is therefore a latency and API result, not a measured recall result. The fair latency comparator is FAISS-Flat-GPU + IDSelectorRange (~0.26 ms, within ~1.3× at N=1M, provisional single-point prototype). Unfiltered exhaustive FAISS Flat retains recall but costs 6.64 ms (33×) and misses the 1 ms AV deadline. The algorithmic move (restricted SGEMV) is small; the contribution is consolidating it with decay, BFS, and streaming append in one API. Native temporal decay folded into the score before top-K (score × exp(-λ·age)). Matches a two-stage FAISS + post-hoc filter pipeline at the same TP@10 = 0.910 in AV perception (0.26 ms vs 0.25 ms p99). Contribution: API consolidation, not raw speedup. Cross-modal bridges in a Neural Shortcut Network (NSN) — warp-cooperative BFS guaranteeing structural reachability of every modality within bounded hops. No claim of semantic optimality; relevance depends on the shared embedding space. At N ≤ 50K, within-modality recall is driven by brute-force cosine + decay; the graph contributes structural cross-modal reachability, not within-modality recall. Hand-fused FP16 cosine (opt-in --use-fp16) wins by 41% at N=10K but loses by 49% at N=1M against cuBLAS Sgemv. Default is cuBLAS; FP16 is a documented small-N opt-in. Default global path: cuBLAS SGEMV + lightweight decay, CUB radix top-K, warp-cooperative BFS over text/audio/image/sensor embeddings in a shared 768-D space. Episode-scoped path: restricted Stage 1 + top-K only. All data GPU-resident; zero per-query allocation. Headline measurements (A100 SXM4 40 GB, D=768) Episode-scoped path: ~200 µs p99 across N=10K…1M (near-flat). Scoped hit@15 = 1.00 is by construction when |episode| < K. Global path: sub-millisecond p99 up to N=50K; 2.67 ms at N=1M; scales to 13M memories (40 GB limit) at 29 ms. On kids-ball, Global hit@15 is 0.79–0.84 vs exhaustive FAISS Flat at 1.0 — the BFS scoring rule plausibly hurts on this contract unless the episode id is supplied. Same-hardware paired-RNG comparison against FAISS-GPU 1.14.1 (Flat + IVF) and cuVS CAGRA 26.04 in results/competitors_20260417/SUMMARY.md. Unfiltered IVF/CAGRA drop to ≤0.4 / 0 hit@15 at N ≥ 10⁵ (known unfiltered-ANN failure mode, quantified on an embodied multimodal contract). Four demonstrators pass empirical p99 deadlines on A100: AV (60 Hz, 1 ms), robot (1 kHz, 1 ms), AR/VR (90 Hz, 5 ms), voice (30 Hz, 20 ms). Episode-scoped at N=1M sits at ~0.20 ms (~80% headroom on the 1 ms AV budget). A raw cuBLAS ring buffer is faster still (0.12 ms at N=2,400) — including on the episode-only case — but does not provide the consolidated contract. All reported latencies are single-query; batched multi-track evaluation is open. Scope and caveats Synthetic corpus. Kids-ball uses correlated 768-D clusters (within-episode cosine > 0.85). Real CLIP/CLAP/E5 geometry differs; absolute hit@15 numbers should not be quoted without a real-encoder re-run. Sequential timestamps (τi = 0.05·i) also confound scope with time (oldest distractors at N=1M demoted ~30–35% by decay). Generator flags --ep-len (for |episode| > K) and --time-mode interleaved are implemented for the next measurement pass. Statistics. p99s over 128–256 paired probes on non-locked-clock vast.ai instances; run-to-run jitter ~±10% at sub-ms scale. The ~1.3× vs IDSelector figure is a single unreplicated prototype. Primary next measurements. (1) |episode| > K so scoped recall is measured; (2) full FAISS+IDSelector / CAGRA-filtered curves across N; (3) interleaved time layout; (4) real-encoder corpus; (5) locked-clock re-runs with ≥10K queries. Source, benchmarks (FAISS-GPU and cuVS CAGRA harnesses, including an IDSelector configuration), diagrams, and LaTeX paper: https://github.com/antonellof/MARS (MIT). CUDA C++17; host-only unit tests; cloud-GPU one-command reproduce. Known issue: optional --use-cuda-graph currently corrupts results between graph replays; global and episode-scoped paths land cleanly without it.