Strategic experiment: pose indexed landmark memory for bounded long horizon world state

Finding

ReWorld, submitted 2026-08-24, separates short horizon control from long horizon memory. At inference it keeps a fixed budget KV cache backed by a pose indexed landmark bank and retrieves landmarks nearest the current pose. The authors report that a fixed 12 chunk cache can revisit the starting view after a 64 second, 384 latent rollout where a sliding window has evicted the evidence and full KV attention runs out of memory.

RuV hypothesis

WorldGraph, RuField, RuView, and RuVector can reuse the memory architecture without reproducing the video generator: retain a bounded hot temporal cache for current control and a spatially indexed landmark bank for durable world evidence.

Proposed primitive

SpatialLandmarkBank<T> keyed by pose or spatial cell with:

• stable landmark identity
• pose and uncertainty covariance
• observation time and freshness
• source provenance root
• content hash
• modality and embedding references
• quality and revisit score
• bounded per region retention policy

Retrieval combines spatial proximity, confidence, freshness, provenance diversity, and semantic relevance. The hot cache stays bounded independently of total exploration time.

Experiment

Use recorded RuView or synthetic spatial trajectories with explicit out and back revisits.

Compare:

1. sliding temporal memory
2. full retained history until memory limit
3. bounded hot cache plus pose indexed landmark bank

Measure revisit accuracy, spatial consistency, memory bytes, lookup latency, stale landmark rate, uncertainty calibration, and robustness to pose noise. Include missing observations and intentionally wrong pose updates.

Security and governance

Landmarks are evidence, not authority. Preserve source provenance and tenancy. A retrieved landmark cannot expand capabilities or overwrite authoritative state without the normal Core Memory or RVM policy path.

Acceptance criteria

At the same hot memory budget, candidate retains at least 2 times longer revisit horizon than sliding memory with less than 10% p95 retrieval latency overhead, no cross tenant leakage, and graceful degradation under injected pose noise.

Reference

arXiv:2608.23565, submitted 2026-08-24.

No implementation files or tests are named. Start by mapping the existing WorldGraph, RuField, RuView, RuVector, Core Memory, and RVM policy paths before designing the experiment around recorded or synthetic trajectories. Done means meeting the stated revisit-horizon, latency, tenant-isolation, and pose-noise degradation criteria.