MemCam: Dual Memory-Enhanced Systems

• MemCam is a dual-faceted approach that enhances memory operations for both video generation with interactive camera control and hardware-accelerated search systems.
• It retains and compresses previous video frames as external memory, using co-visibility scoring to maintain scene consistency over long trajectories.
• It employs memristor-based content-addressable memory with configurable parameters for rapid point and range search, balancing energy, latency, and endurance.

MemCam denotes distinct approaches in two technical domains: (1) memory-augmented neural architectures for consistent video generation with dynamic camera control (Gao et al., 27 Mar 2026), and (2) memristor-based configurable content-addressable memory systems for accelerating search in storage-class memory (Liu et al., 2016). Both share a core emphasis on augmenting memory access and retrieval to improve performance—scene consistency in the first, and search efficiency in the second. The following presents a comprehensive technical exposition of both MemCam paradigms.

1. Memory-Augmented Camera Control for Video Generation

MemCam (Gao et al., 27 Mar 2026) addresses the challenge of interactive video generation where, at each time step $t$, the camera pose $(\mathbf{R}_t, \mathbf{t}_t)$ is user-specified, and the system must synthesize a frame $F_t$ consistent with the new view and preceding content. Naïve video diffusion methods, when confronted with long sequences and large camera trajectories, tend to lose scene consistency, manifesting as drift, hallucinated content, or geometric inconsistencies.

MemCam introduces a memory-augmented video diffusion transformer that retains all previously generated frames as external memory and leverages contextual encoding and retrieval to maintain long-term scene alignment. The workflow incorporates:

• Storing the set $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$ as explicit external memory.
• Compressing each memory frame $F_i$ via an encoder $E$ to obtain $M_i = E(F_i)$; $$M_i \in \mathbb{R}^{\frac{H}{2}\times\frac{W}{2}\times C}$$ where $C$ denotes model channel width.
• Computing a relevance score $s(i,t)$ via co-visibility IoU. For $(\mathbf{R}_t, \mathbf{t}_t)$0 sampled 3D points $(\mathbf{R}_t, \mathbf{t}_t)$1,

$(\mathbf{R}_t, \mathbf{t}_t)$2

where $(\mathbf{R}_t, \mathbf{t}_t)$3 indicates whether $(\mathbf{R}_t, \mathbf{t}_t)$4 is inside the camera’s view frustum at time $(\mathbf{R}_t, \mathbf{t}_t)$5.

• Selecting the top-$(\mathbf{R}_t, \mathbf{t}_t)$6 most relevant past frames based on $(\mathbf{R}_t, \mathbf{t}_t)$7 and fusing their compressed tokens together with the current noisy prediction tokens $(\mathbf{R}_t, \mathbf{t}_t)$8:

$(\mathbf{R}_t, \mathbf{t}_t)$9

with $F_t$0, $F_t$1.

• Feeding $F_t$2 through DiT self-attention layers and decoding the result with a 3D-VAE head.

This approach ensures that context relevant to the current camera trajectory, regardless of temporal distance, is efficiently available for each generative prediction, directly addressing the main bottleneck of scene consistency in interactive video synthesis.

2. Loss Functions and Training Protocol

MemCam is trained end-to-end with a mean-squared-error denoising objective typical for diffusion models: $F_t$3 where $F_t$4; the context includes both the camera pose embedding and the compressed memory latents.

No auxiliary adversarial objectives or explicit temporal-consistency losses are used. For robustness, a regularization step zero-masks all selected memory latents in 10% of training steps. Additionally, the first frame is always included in context, and real camera trajectories are sampled during training to promote smooth transitions.

3. Efficiency and Computational Complexity

The computational costs in MemCam arise primarily from memory retrieval and attention integration:

• Co-visibility retrieval: Each $F_t$5 requires $F_t$6 point-visibility tests across $F_t$7 memory frames, for total $F_t$8. With $F_t$9, the costs are mitigated by parallel GPU processing.
• Token compression: Each memory frame, without compression, yields $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$0 tokens. Compression results in $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$1 tokens per frame. Thus, attention cost per forward pass is reduced by $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$2 per memory frame.
• Empirical speedup: MemCam-76 achieves approximately $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$3 speedup with $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$4-frame compressed memory relative to an uncompressed baseline, without sacrificing reconstruction quality.

4. Inference Pipeline

The inference procedure for MemCam proceeds as follows:

$E$5

This reflects the explicit memory selection, compression, and fusion steps integral to MemCam (Gao et al., 27 Mar 2026).

5. Empirical Performance and Ablation Studies

MemCam demonstrates significant gains over baseline and state-of-the-art methods on both 90° and 360° round-trip video generation benchmarks, using the Context-as-Memory (100 scenes, 7,901 frames/scene) and RealEstate10K datasets. Key metrics (higher is better for PSNR, SSIM; lower is better for FVD):

Method	PSNR (360°)	SSIM (360°)	FVD (360°)
MemCam	14.81	0.423	167.9
GF	10.07	0.402	852
DFoT	8.94	-	1,188

Ablation confirms both co-visibility selection and context compression are indispensable: compressed memory ("Ours-76") matches quality of uncompressed ("None-76") at $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$5 speedup.

MemCam's improvements are most pronounced for extended camera trajectories involving large viewpoint changes and revisits, which cause catastrophic forgetting or artifacts for naїve temporal conditioning schemes.

6. Memristor-Based MemCAM for Search Acceleration

Independent of video generation, MemCAM also refers to a configurable memristor-based content-addressable memory system (Liu et al., 2016), designed to accelerate point and range search queries in storage-class memory. MemCAM employs a crossbar substrate where each two-terminal memristor at the row-column intersection may act as storage (under $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$6 voltage) or as a logic gate (under $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$7 or $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$8 voltages).

Circuit Operation and Search

• Each cell encodes data and key bits in a 2-bit one-hot fashion with supplementary memristors for comparison and storage.
• An 11-step implication-logic sequence supports equality, less-than, and greater-than comparisons; a 10-step reduction across columns yields an Entry Match Signal (EMS).
• Point search returns a row only if EMS=1; range search leverages an additional comparison step.
• All $$\mathcal{M} = \{(F_i, \mathbf{R}_i, \mathbf{t}_i)\}_{i=1}^{t-1}$$9 entries participate in parallel; latency is determined by key length and wiring.

Performance and Endurance Characteristics

• Energy per bit per search: $F_i$0 fJ/bit/search.
• Search latency: $F_i$1 ns.
• Endurance: $F_i$2 writes per device, resulting in lifetimes from $F_i$3 years (T-tree-CAM) to $F_i$4 years (TB$F_i$5-tree) for $F_i$6 search/s, governed by search structure and write frequency.

Hybrid CMOS–MemCAM Structures

Partitioning logical search trees across fast, durable CMOS memory and parallel, less-enduring memristor arrays enables tunable trade-offs:

• Hash-CAM: Hash-based partition in CMOS, each mapped to one CAM partition.
• T-tree-CAM / TB$F_i$7-tree / TB$F_i$8-tree-CAM: Orders and depths control fan-out, capacity, and endurance-latency trade-offs.
• CAM partition size, number, and B$F_i$9-tree branching order are explicit knobs for performance-lifetime tuning.

7. Configurability and Design Trade-offs

In memristor-based MemCAM, several configurable parameters determine system behavior:

• Voltage levels ($E$0, $E$1, $E$2) control storage versus compute mode.
• Array size and CAM/B$E$3-tree partitioning impact capacity, parallelism, and device wear.
• Logical tree depth, fan-out, and data rotation heuristics provide fine-grained control over endurance versus performance.
• Key width $E$4 directly sets search step count and associated energy/latency.

Selecting appropriate configuration leverages analytic models for energy, latency, and expected device lifetime, with design space spanning from ultra-low-latency architectures (minute-scale lifetime) to robust, decade-long solutions (Liu et al., 2016).

---

MemCam thus encompasses advanced architectures for both neural video generation with long-horizon scene consistency and high-performance storage systems with configurable latency-endurance profiles. Each variant exemplifies the exploitation of external memory references—either neural or hardware—to mitigate context loss or search inefficiency, achieving significant advances within their respective domains (Gao et al., 27 Mar 2026, Liu et al., 2016).