What Spatial Memory Must Store: Occlusion as the Test for Language-Agent Memory

# Occlusion as the Test for Language-Agent Memory
Source: [https://arxiv.org/html/2606.10299](https://arxiv.org/html/2606.10299)
## What Spatial Memory Must Store: Occlusion as the Test for Language\-Agent Memory

\(June 2026\)

###### Abstract

Language\-agent “memory palace” systems anchor each memory to a world coordinate, on the intuition that geometry adds something text cannot\. We make that intuition testable and report three results\. First, the memory\-palace default of folding spatial proximity into a linear blend beside recency and importance does not help and can hurt: in a pre\-registered recall experiment the shipped blend fails its own frozen test \(meanΔ\\DeltaHit@5−0\.0375\-0\.0375, Wilcoxonp=0\.306p=0\.306, CI including0\), sitting at a position\-blind baseline, while a geometry\-led weighting wins decisively \(\+0\.3208\+0\.3208,p<10−15p<10^\{\-15\}\): geometry must*lead*recall when the query regime is spatial\. We demonstrate the spatial\-dominant ranking form; the stronger index/gate form remains future work\. Second, memory*recall*and*visibility*must be separated: recall is occlusion\-blind by design \(you correctly remember the next room behind a wall\), while visibility is a perception predicate over stored geometry that the live system never computed\. A one\-line ray\-versus\-voxel digital differential analyzer \(DDA\), re\-pointed from the gaze ray the agent already casts, supplies it: text and the live FoV cone both score0\.0000\.000on849849behind\-wall targets while cone\-plus\-DDA reaches0\.9820\.982\(exact McNemarp<10−6p<10^\{\-6\}\); coordinate recall separately resolves near\-duplicate locations a cosine null cannot \(1\.0001\.000versus0\.5330\.533,n=150n\{=\}150\)\. Third, the visibility predicate is confirmed live under a git\-committed pre\-registration \(SPMEM\-OCC\-LIVE\-v1: eight scripted worlds of one jittered occluder family, automated oracle scoring,9696behind\-wall targets, false\-visible1\.000→0\.0001\.000\{\\to\}0\.000, pooled exact McNemarp=2\.5×10−29p\{=\}2\.5\{\\times\}10^\{\-29\}\), a run that surfaced and fixed a real relay anchor defect\. We concede that occlusion\-needs\-geometry is near\-tautological; the contribution is the measurement and isolation, separating what spatial memory must*store*from how it is read\. A preliminary action\-level ablation \(an effect of object binding, not geometry per se\) lifts situated\-action accuracy from0\.6250\.625to1\.0001\.000\(p=0\.0039p=0\.0039\)\. These pilots power a frozen confirmatory study \(SPMEM\-ZERO\-REAL\-PREREG\-v1\); the full human\-authored multi\-world study with blind raters remains future work\.

## 1Introduction

A growing family of language\-agent systems gives each agent a persistent memory of what it observed, did, and said, and retrieves from that memory to plan the next action\[[1](https://arxiv.org/html/2606.10299#bib.bib1),[2](https://arxiv.org/html/2606.10299#bib.bib2)\]\. A distinctive subfamily, inspired by the method of loci and by embodied 3D scene memory\[[6](https://arxiv.org/html/2606.10299#bib.bib6),[7](https://arxiv.org/html/2606.10299#bib.bib7)\], anchors each memory to a*world coordinate*: the agent stores not only*what*happened but*where*, and later retrieves memories by location and visibility\. The appeal is that place becomes a free retrieval cue\. We built such a system, a shared voxel world in which external language\-model “brains” act through a renderless tool interface and anchor each memory to a 3D cell \(Figure[1](https://arxiv.org/html/2606.10299#S1.F1)\), and we ask the question the appeal demands: what, precisely, does the geometry buy that text does not already provide?

![Refer to caption](https://arxiv.org/html/2606.10299v1/figures/fig_world_village.png)Figure 1:The Zero world the memory system lives in: an external\-brain agent society acts in and builds persistent structures in a shared voxel 3D space \(an agent\-built village shown, captured from the live web client\)\. Agent memories are subject\-anchored to locations in this world\. This paper asks what geometry such memories must read to do work that text cannot\.We are not first to recover occluded content from stored geometry\. Render\-as\-recall systems, GSMem\[[8](https://arxiv.org/html/2606.10299#bib.bib8)\]\(3D Gaussian\-splat memory rendered from unoccluded viewpoints\) and RenderMem\[[9](https://arxiv.org/html/2606.10299#bib.bib9)\]\(rendering from query\-implied viewpoints to reason about visibility and occlusion\), demonstrate that insight before us, and we claim no priority on it\. Our contribution is upstream of rendering: we isolate what must be stored from how it is read, showing that once geometry is stored, a one\-line ray\-march and an LLM reading the coordinates as text recover occlusion at parity \(E3\), so the irreducible requirement is geometry storage, not a renderer; and we confirm the perception asymmetry live under pre\-registration on a running text\-agent system rather than only in simulation\.

Our framing follows the embodied\-memory literature\. 3D\-Mem\[[6](https://arxiv.org/html/2606.10299#bib.bib6)\]and 3DSPMR\[[7](https://arxiv.org/html/2606.10299#bib.bib7)\]treat spatial memory as capturing places visually, remembering object locations, occlusion, and viewpoint, and recalling by location and visibility, with field\-of\-view coverage and visibility used as gating and verification signals rather than as a ranking axis\. Read this way, the test of a spatial memory is sharp:if a non\-spatial text or vector index could answer the query, it is not a spatial\-memory test\.Geometry must do work text cannot, occlusion being the cleanest case\. A geometry\-blind index cannot derive “behind” from captions alone: occlusion requires the coordinates, the occluders, and a line\-of\-sight computation \(handed the coordinates, even a text\-medium reasoner recovers it, Section[7](https://arxiv.org/html/2606.10299#S7)\)\.

It helps to separate two distinct operations the literature can blur\. Memory*recall*retrieves by*location*: anchored to a coordinate, it correctly surfaces what is stored at or near a place, including the next room behind a wall, so recall is occlusion\-blind*by design*\.*Visibility*is a different thing: a perception verification \(“is this target visible from where I stand”\), a queryable geometric predicate a recall result can be asked to answer\. Occlusion is therefore a question memory must be able to*answer*, not a suppressor that hides behind\-wall memories\. Our test measures that perception predicate; recall itself keeps returning the behind\-wall memory\.

This reframing also corrects our own earlier work\. A prior program organized a document corpus into topic “buildings” and recalled by routing a query to the right building, and reported a “spatial” win\. An audit found it studied topic\-partition text retrieval: the ground truth was a refinement of the partition under test, the retrieval never read the 3D coordinates, and two of three headline wins do not survive a family\-wise correction\. We present that result honestly in Section[4](https://arxiv.org/html/2606.10299#S4)as a scoping negative, then build the real test on the one asymmetry Zero gives us cleanly: no visibility query exists anywhere in the live perception path, yet a line\-of\-sight answer is exactly the geometry a geometry\-blind text index cannot compute\.

#### Contributions\.

1. 1\.A definition and a test\.Spatial memory must store the geometry that its predicates need, visibility, occlusion, containment, and viewpoint, and the test of a spatial\-memory claim is whether the query requires geometry a content\-only index cannot compute\. We give a minimum\-representation schema \(what to store, what to compute\) per query type \(Table[1](https://arxiv.org/html/2606.10299#S3.T1)\)\.
2. 2\.Perception evidence\.Coordinate recall solves near\-duplicate localization that a text null cannot \(1\.0001\.000versus0\.5330\.533\); field\-of\-view\-only*perception*cannot tell a visible target from a behind\-wall one; a ray\-versus\-voxel DDA line\-of\-sight predicate answers it in a controlled voxel simulation \(0\.9820\.982\) and is confirmed on the live relay under a git\-committed pre\-registration \(SPMEM\-OCC\-LIVE\-v1: eight worlds,9696behind\-wall targets, false\-visible0\.0000\.000versus1\.0001\.000for a cosine null and the FoV cone, pooled exact McNemarp=2\.5×10−29p\{=\}2\.5\{\\times\}10^\{\-29\}\), a run that first surfaced and fixed a relay bug that had marched the predicate to the wrong anchor\.
3. 3\.Retrieval\-mechanism evidence\.A pre\-registered recall experiment shows the shipped diluted linear blend fails its own frozen test \(it ties or loses to a position\-blind baseline\) while geometry\-led \(spatial\-dominant\) weighting wins decisively, making the index\-versus\-ranker distinction quantitative\.
4. 4\.Boundary evidence\.Topic\-partition retrieval and object\-binding action gains are useful but*not*geometry\-specific, and a set of robustness checks \(E1–E5\) narrows the claim to where geometry is irreducible\.

We are explicit that 3DSPMR\[[7](https://arxiv.org/html/2606.10299#bib.bib7)\], not this paper, is first to use field\-of\-view/visibility geometry as a prior in spatial agent memory; we claim no priority on that idea\. What is new here is the measured ranker\-versus\-predicate distinction, the storage\-not\-medium isolation \(E3\), and the live pre\-registered occlusion confirmation \(SPMEM\-OCC\-LIVE\-v1\), none of which 3DSPMR or the render\-as\-recall systems\[[8](https://arxiv.org/html/2606.10299#bib.bib8),[9](https://arxiv.org/html/2606.10299#bib.bib9)\]run\.

We make no head\-to\-head, SOTA\-beating claim: we run no benchmark in which our system outperforms render\-as\-recall \(GSMem\[[8](https://arxiv.org/html/2606.10299#bib.bib8)\], RenderMem\[[9](https://arxiv.org/html/2606.10299#bib.bib9)\]\) or FoV\-prior \(3DSPMR\[[7](https://arxiv.org/html/2606.10299#bib.bib7)\]\) systems on a shared task\. Our contribution is isolation and measurement, separating what spatial memory must store from how it is read, and a pre\-registered live confirmation, not a leaderboard win\. A head\-to\-head program against those neighbors on a shared occlusion query set is the central piece of future work \(Section[9](https://arxiv.org/html/2606.10299#S9)\)\.

The full human\-authored multi\-world confirmatory study is pre\-registered \(SPMEM\-ZERO\-REAL\-PREREG\-v1\); its occlusion slice has now run live under its own freeze \(SPMEM\-OCC\-LIVE\-v1, Section[6\.9](https://arxiv.org/html/2606.10299#S6.SS9)\), and the pilots here power and de\-risk the rest\.

## 2Related Work

#### Agent memory\.

Generative Agents\[[1](https://arxiv.org/html/2606.10299#bib.bib1)\]score memories by a weighted sum of recency, importance, and relevance; this remains the canonical text\-memory baseline\. MemGPT\[[2](https://arxiv.org/html/2606.10299#bib.bib2)\]\(now Letta\) treats the context window like an operating\-system memory hierarchy, paging between a working context and an archival store whose retrieval is vector search\. A recent wave structures memory as a graph: A\-MEM\[[3](https://arxiv.org/html/2606.10299#bib.bib3)\]builds an evolving Zettelkasten of linked notes; Mem0\[[4](https://arxiv.org/html/2606.10299#bib.bib4)\]extracts and consolidates facts with an optional graph variant; Zep/Graphiti\[[5](https://arxiv.org/html/2606.10299#bib.bib5)\]maintains a temporal knowledge graph\. These methods are powerful, but they introduce a graph that must be*extracted*by a language model, with attendant cost and drift\. A spatial anchor is an extraction\-free structural prior of a different kind, and the question of this paper is which part of it is doing real work\.

#### Spatial and embodied memory\.

3D\-Mem\[[6](https://arxiv.org/html/2606.10299#bib.bib6)\]represents an explored scene with multi\-view “memory snapshots” for embodied reasoning\. “Vision to Geometry” \(3DSPMR\)\[[7](https://arxiv.org/html/2606.10299#bib.bib7)\]uses field\-of\-view coverage as an explicit geometric prior, applying it for memory gating, for verifying whether enough has been seen before answering, and as an exploration incentive; its ablation shows these FoV mechanisms \(not a ranking signal\) drive the gains\. We concede this precedent plainly: 3DSPMR is the nearest prior work and FoV\-geometry\-as\-prior is not new here\. Two things it does not do define our delta\. It does not empirically separate geometry as a verification predicate from geometry folded into a blended ranking term, the distinction our pre\-registered recall experiment measures \(Section[6\.8](https://arxiv.org/html/2606.10299#S6.SS8)\), where blending dilutes below a position\-blind baseline while a geometry\-led weighting wins; the verification predicate itself is measured separately by the occlusion pilots\. And it does not decouple occlusion\-blind recall from a separate visibility query the geometry answers, which our live test isolates\. Our differentiator is that measured ranker\-versus\-predicate distinction and the recall/perception split, not the use of visibility geometry itself\. These works are the rigor we emulate: ground truth authored from the world’s geometry, value measured by a downstream\-task delta and a geometry\-on/geometry\-off ablation, and an accuracy metric paired with an efficiency metric\. They motivate our central claim: geometry is an index and verification mechanism, occlusion and visibility being the irreducible case\. Large\-scale agent societies such as PIANO/Project Sid\[[10](https://arxiv.org/html/2606.10299#bib.bib10)\]are the honest null baseline, rich behavior from non\-spatial text memory\.

#### Render\-as\-recall\.

GSMem\[[8](https://arxiv.org/html/2606.10299#bib.bib8)\]stores a 3D Gaussian\-splat representation of an explored environment and recovers occluded content by rendering novel views from unoccluded viewpoints\. RenderMem\[[9](https://arxiv.org/html/2606.10299#bib.bib9)\]renders from query\-implied viewpoints to reason explicitly about visibility and occlusion\. Both demonstrate, before this paper, that storing geometry enables occlusion recovery; we claim no priority on that insight\. Three things distinguish our approach\. First, our geometry is authoritative voxel occupancy \(an oracle\) rather than a 3DGS reconstruction with potential surface error\. Second, E3 shows that text coordinates recover occlusion at parity with ray\-marching \(0\.99 versus 0\.985, McNemarp=1\.0p=1\.0\), so storage and not a renderer is the irreducible requirement; neither GSMem nor RenderMem separates those two\. Third, we confirm the occlusion asymmetry live under a git\-committed pre\-registration on a running text\-agent system, rather than in simulation only\.

#### World models and the missing memory layer\.

The geometry\-free memory systems above \(Generative Agents, MemGPT, A\-MEM, Mem0, Zep\) are the concrete evidence for the gap: each is deployed or peer\-reviewed and not one stores spatial geometry or computes a visibility predicate, so none can answer “is this remembered thing visible from here\.” That absence is the missing layer\. A parallel program builds*world models*: omnimodal systems that render, simulate, and plan over physical or virtual scenes \(e\.g\. NVIDIA Cosmos\[[16](https://arxiv.org/html/2606.10299#bib.bib16)\]\)\. A functional taxonomy of that program\[[17](https://arxiv.org/html/2606.10299#bib.bib17)\]usefully frames it as three projections of the same agent–action–state–observation loop: a*renderer*outputs observations \(pixels\), a*simulator*outputs faithful state \(geometry, physics\), and a*planner*outputs actions\. We cite the taxonomy as framing, not evidence: the load\-bearing fact is the measured geometry\-blindness of the peer\-reviewed memory stack \(Section[6\.5](https://arxiv.org/html/2606.10299#S6.SS5)\)\. All three world\-model projections concern the*instantaneous*loop, what the world looks like, how it would evolve, what to do next\. None of them specify what a long\-horizon agent must*retain across*that loop: which places, objects, occlusions, and viewpoints persist, why a memory is anchored where it is, and how a human or agent later inspects and corrects it\. That persistence\-and\-inspection layer is world*memory*, and it is the subject of this paper\. Our claim is complementary to the world\-model program rather than competing with it: world models generate and predict worlds; spatial memory stores the geometry\-dependent facts \(visibility, occlusion, containment, viewpoint\) those worlds leave behind, so agents can act over time and people can audit the result\. A renderer or simulator could in principle*supply*the occluder geometry our predicate reads, but neither, nor a content\-only text memory, decides*what to keep*or*answers*“is this remembered thing visible from here\.”

#### Retrieval\.

BM25\[[14](https://arxiv.org/html/2606.10299#bib.bib14)\]remains a strong lexical baseline; dense bi\-encoder retrieval and their reciprocal\-rank fusion\[[15](https://arxiv.org/html/2606.10299#bib.bib15)\]are the production standard\. RAPTOR\[[11](https://arxiv.org/html/2606.10299#bib.bib11)\]recursively embeds, clusters, and summarizes a corpus into a tree; its cluster\-then\-retrieve core is the closest non\-spatial analog to routing a query into a region\. GraphRAG\[[12](https://arxiv.org/html/2606.10299#bib.bib12)\]builds an entity graph with community summaries for global, query\-focused summarization\. HippoRAG\[[13](https://arxiv.org/html/2606.10299#bib.bib13)\]runs personalized PageRank over an extracted knowledge graph, inspired by hippocampal indexing\. These are the comparators for the text\-retrieval regime; none of them can answer an occlusion query, which is the point\.

## 3The Index\-versus\-Ranker Distinction

Before the experiments, we fix what “store the geometry” means per query type\. The title asks what spatial memory must*store*; Table[1](https://arxiv.org/html/2606.10299#S3.T1)answers it as a minimum\-representation schema, pairing each spatial query type with the data the memory must hold, the predicate recall must compute over it, and the way a content\-only memory fails\. Every row is a query whose answer is a function of coordinates and occupancy, not of the stored text, which is exactly why a geometry\-blind index cannot serve it\.

Table 1:Minimum\-representation schema: what spatial memory must*store*and*compute*per query type, and how a content\-only \(text or vector\) memory fails\. Each predicate is a function of geometry the stored content does not encode; storing the geometry is the irreducible step, after which a cheap computation recovers the answer\.Let a memorymmhave an embeddingeme\_\{m\}and an anchorxm∈ℝ3x\_\{m\}\\in\\mathbb\{R\}^\{3\}\. A query has an embeddingqqand a standpointc∈ℝ3c\\in\\mathbb\{R\}^\{3\}with a facing direction\. There are two ways to use the anchor\.

#### Spatial as ranker\.

Score every memory by a blend of semantic similarity and spatial proximity,

s​\(m\)=wr​cos⁡\(q,em\)\+ws​ϕ​\(∥c−xm∥\)\+…,s\(m\)=w\_\{r\}\\,\\cos\(q,e\_\{m\}\)\+w\_\{s\}\\,\\phi\\\!\\left\(\\lVert c\-x\_\{m\}\\rVert\\right\)\+\\dots,and rank byss\. This is the memory\-palace default\. It can only help when, for the true targets, small∥c−xm∥\\lVert c\-x\_\{m\}\\rVertcoincides with high relevance, a property of where things were placed and where the agent recalls from, not of the retriever\.

#### Spatial as index and verification\.

Use the anchor to*select*a candidate set, then verify a geometric predicate the score never encodes\. The predicate may be containment \(which region is the query in\), a field\-of\-view cone \(is the target in front of the observer\), or, the irreducible case,*visibility*: is there a clear line of sight from observer to target, or is the target behind a wall? Here geometry never enters a distance score\. It selects and it verifies, and the verification is something text cannot compute\.

The distinction matters because the two roles fail differently\. A spatial ranker bets about world layout and can help or hurt with placement\. A visibility predicate makes no bet: occlusion is a fact of the world’s geometry, true regardless of how memories are placed or embedded\. The ranker\-blend failure mode argued here, that folding spatial proximity into a linear score beside recency and importance can dilute the spatial signal below a position\-blind baseline, is not hypothetical: it is measured directly in a pre\-registered recall experiment \(Section[6\.8](https://arxiv.org/html/2606.10299#S6.SS8)\), where the shipped blend ties or loses to pure vector while a geometry\-led \(spatial\-dominant\) weighting wins decisively\.

We are careful about what that win is\. A spatial\-dominant weighting is still a*ranker*: it raises spatial to the leading term of a linear score but never cuts the candidate set by geometry first\. A true*index*or*gate*, select by a geometric predicate \(containment, line of sight\) and only then rank, is a stronger form of the same idea that we did*not*run here; it is future work\. Stating it as a calibrated thesis: spatial structure should not be a small diluted term in a generic linear blend\. It must*lead*retrieval, whether as an index/gate, a verification predicate, or a spatial\-dominant score when the query regime is spatial\. Our pre\-registered experiment demonstrates the spatial\-dominant \(geometry\-led ranking\) form; the index/gate form is stronger and left to future work\. The demonstrated fix in Section[6\.8](https://arxiv.org/html/2606.10299#S6.SS8)is therefore the geometry\-led\-ranking endpoint of this distinction, not yet the pure index\. The rest of the paper measures the verification role on the two cleanest cases\.

## 4What does not count: topic\-partition retrieval is not spatial memory

We report this honestly because it shaped the design and because hiding it would misrepresent what we know\. A prior program built an organized library: a 240\-document technical corpus, 40 documents from each of six topics, with each topic placed in its own “building” on a3×23\\times 2grid, and recall by routing a query to the nearest building centroid before ranking inside it\. On 142 leave\-one\-out queries with citation\-graph ground truth and real 384\-dimensional embeddings, route\-then\-rank matched flat vector search at roughly a sixth of the candidate cost, and with a hybrid in\-region ranker reached the highest recall@10 among the fair methods tested \(0\.3930\.393versus0\.3370\.337dense, at3939candidates instead of239239\)\.

A subsequent audit found that this is*not*a spatial\-memory result, for three reasons, each fatal to the spatial claim\.

- •The ground truth refined the partition under test\.Cross\-region relevant fraction was0\.0000\.000by construction: every citation\-related document lived in the same building\. The partition the retriever used was the partition that scored it, a tautology\.
- •The retrieval never read coordinates\.Every run used spatial weightpw​\_​spatial=0p\_\{w\}\\\_\\mathrm\{spatial\}=0, and the document anchor was never dereferenced\. The “spatial” axis contributed literally nothing; the discriminating signal was always lexical and semantic topic\.
- •The headline wins do not survive correction\.The two significant deltas \(versus dense RAG atp=0\.018p=0\.018and versus hybrid RRF atp=0\.045p=0\.045\) do not clear a Holm\-Bonferroni family\-wise threshold; only the gain over k\-means routing \(p=0\.0005p=0\.0005\) survives\.

What that study actually showed is real but modest, and we keep it as such: organizing a corpus by topic and routing to the right partition is a cheap candidate\-narrowing trick that does not lose accuracy\. That is a document\-library retrieval finding, not a demonstration that geometry does work text cannot\. The honest verdict is a scoping negative:*a method that never reads coordinates and is scored by a refinement of its own partition cannot be evidence about spatial memory\.*It told us what does not count, and pointed us at what does\.

## 5Method

### 5\.1Embodied capture

The agent forms memory through its own perception and movement loop, using existing Zero tools\. It enters a space, walks a coverage\-driven path, and at each stop captures an observation:look\_aroundreturns its pose, yaw and pitch, and alooked\_at\_subject, the world center of the grid cell the gaze ray hits\. The gaze ray is a DDA over the voxel field that reads real geometry through theisSolidAtprimitive\. For each salient gaze hit the agent writesappend\_memory\(content, position=body, subjectPosition=looked\_at\_subject\), the relay’s subject\-anchor path: the memory is bound to the thing it is about \(the cell observed\), not to where the agent happened to stand\. We use the grid\-to\-world conventionwx=0\.5​gx,wy=2\.0\+0\.5​gy,wz=0\.5​gzw\_\{x\}=0\.5\\,g\_\{x\},\\;w\_\{y\}=2\.0\+0\.5\\,g\_\{y\},\\;w\_\{z\}=0\.5\\,g\_\{z\}throughout\.

### 5\.2Memory representations

All representations are distilled from the same captured observation stream; they differ only in what they store and how recall reads it\.

- •M0 \(text null\)\.Each observation becomes a natural\-language sentence, embedded and retrieved by pure cosinekk\-NN\. No coordinate is read at recall, by construction\. If M0 answers a query, the query is not a spatial test\.
- •M1b \(coordinate recall\)\.The live relay radius recall: it reads the memory anchor and the query center and ranks by 3D Euclidean distance \(spatial weight\>0\>0\)\. No field of view, no occlusion\.
- •M3a \(FoV\-cone, live\)\.The live field\-of\-view*visibility\-perception*arm, which answers “is targetTTvisible from observerPP”\. It derives a forward vector from the observer facing and prefilters candidates by a dot\-product cone \(the cone normal is\[sin⁡ry,0,cos⁡ry\]\[\\sin r\_\{y\},0,\\cos r\_\{y\}\]\)\. It tests whether a target is in front of the observer\.It contains no line\-of\-sight test:a behind\-wall target inside the cone is judged visible exactly like an open one, so the perception answer is wrong\.
- •M3\-occ \(cone plus line\-of\-sight\)\.The same visibility\-perception query, M3a plus the occlusion predicate\. For observerPPand a candidate targetTTinside the cone, cast a DDA rayP→TP\\to Tstepping one cell at a time and testisSolidAtat each intermediate cell\.TTis occluded if any intermediate solid cell lies strictly betweenPPandTT, visible otherwise\. This reuses the same ray\-versus\-voxel primitive the gaze ray already runs, re\-pointed from observer\-to\-gaze\-cell to observer\-to\-target\. M3a and M3\-occ do not drop or rank memories; they answer the visibility question over a recalled target\.

M2 \(object\-at\-location scene graph\) and M4 \(rendered ego\-view with a VLM reader\) require infrastructure that does not exist in the live system today and are deferred to a Tier\-B future study \(Section[9](https://arxiv.org/html/2606.10299#S9)\)\.

### 5\.3The occlusion DDA

The fix is small and reuses what exists\. Zero’s gaze ray already steps a DDA through the voxel field callingisSolidAtto find what the agent is looking at\. The occlusion predicate is the same loop re\-pointed: instead of casting from the observer along the gaze direction until a solid cell, cast from the observer toward each candidate target and report whether any solid cell intervenes\. No new geometry, no new data structure, one re\-used primitive\. The live perception path has no visibility query that runs it; that absence is precisely what M3a’s failure measures\.

## 6Results \(Pilots\)

We report a suite of controlled pilots, each labeled as a pilot\. The first runs against the live relay recall RPC; the second runs in a controlled JavaScript voxel simulation that mirrors Zero’sisSolidAtgeometry\. They isolate the asymmetry on the trivial and the hard case respectively\. They are not the full study; they power and de\-risk the pre\-registered confirmatory design \(Section[10](https://arxiv.org/html/2606.10299#S10.SS0.SSS0.Px1)\)\.

### 6\.1Pilot 1: near\-duplicate localization \(the trivial case\)

Two structures with identical content are distinguishable only by where they are\. Vector embeddings of identical text collide, so a text null cannot tell them apart; coordinate recall can\. We ran 150 trials on the live relay \(space 1163, agent 1129\): each trial asks which of two near\-duplicate memories is the one at a queried location\. M1b \(coordinate recall, spatial weight22with a query center set on all150/150150/150trials\) is compared against M0 \(the pure\-cosine null, vector\-only\)\.

Table 2:Pilot 1, near\-duplicate localization on the live relay recall RPC\. M1b reads coordinates and resolves identical\-content memories by location; M0 cannot\. Exact McNemar over discordant pairs \(7070M1b\-right/M0\-wrong,0M0\-right/M1b\-wrong\) givesp<10−6p<10^\{\-6\}\.Coordinate recall is perfect \(1\.0001\.000\) while the text null sits near chance \(0\.5330\.533\)\. The discordant pairs are entirely one\-sided:7070trials where M1b is right and M0 is wrong,0the other way, so the exact McNemar test givesp<10−6p<10^\{\-6\}\(Table[2](https://arxiv.org/html/2606.10299#S6.T2), Figure[2](https://arxiv.org/html/2606.10299#S6.F2)\)\. The geometry probe confirms M1b dereferenced coordinates on all150150trials \(spatial weight22, query center set\), while M0 stayed vector\-only\. This is the easy half of the asymmetry: where content collides, location resolves it, and text cannot\.

![Refer to caption](https://arxiv.org/html/2606.10299v1/x1.png)Figure 2:Pilot 1 \(live relay\), near\-duplicate localization overn=150n=150trials\. Coordinate recall \(M1b\) reaches1\.0001\.000accuracy versus0\.5330\.533for the pure\-cosine text null \(M0\)\. The discordant pairs are one\-sided \(7070versus0\), exact McNemarp<10−6p<10^\{\-6\}\. Where two memories share identical content, only location distinguishes them\.
### 6\.2Pilot 2: occlusion / line\-of\-sight \(the hard case\)

This is the case a geometry\-blind text index cannot compute\. We built a controlled voxel world: a wall slab spanningx∈\[9\.5,10\.5\]x\\in\[9\.5,10\.5\],z∈\[0,20\]z\\in\[0,20\],y∈\[0,10\]y\\in\[0,10\]with a doorway gap atz∈\[9,11\]z\\in\[9,11\]\. From a fixed observer we enumerated11441144targets inside the field\-of\-view cone, of which849849are occluded by the wall and295295are genuinely visible \(through open air or the doorway\)\. Ground truth is the world’s geometry, computed independently of any memory representation by fine0\.050\.05m analytic segment sampling\. We compare three arms: M0 \(text\), M3a \(the FoV cone, faithfully reimplementing the live recall’s cone logic in this simulation\), and M3\-occ \(cone plus the DDA line\-of\-sight\), where M3\-occ uses a deliberately*coarse*0\.50\.5m DDA, a different implementation of the same physical visibility fact, so the comparison is not circular\.

Table 3:Pilot 2, occlusion in a controlled voxel world \(wall slab with a doorway;11441144in\-cone targets,849849occluded,295295visible\)\. Text and the FoV cone call every in\-cone target visible, so they are0%0\\%correct on behind\-wall targets \(100%100\\%false\-visible\)\. Adding the DDA line\-of\-sight primitive recovers occlusion\. Exact McNemar M3\-occ versus M3a on the occluded subset:834834discordant in M3\-occ’s favor,0the other way,p<10−6p<10^\{\-6\}\. Ground truth is fine0\.050\.05m analytic sampling; M3\-occ is a coarse0\.50\.5m DDA \(non\-circular\)\.The result is unambiguous \(Table[3](https://arxiv.org/html/2606.10299#S6.T3), Figure[3](https://arxiv.org/html/2606.10299#S6.F3)\)\. M0 and M3a are*identical*on this task at0\.2580\.258overall and0\.0000\.000on the occluded subset: both call every in\-cone target visible, so both are wrong on every single one of the849849behind\-wall targets, a100%100\\%false\-visible rate\. The live FoV cone adds nothing over text here, because being in front of the observer says nothing about being seen\. M3\-occ reaches0\.9870\.987overall and0\.9820\.982on the occluded subset, with a0\.0180\.018false\-visible rate\. The exact McNemar test between M3\-occ and M3a on the occluded subset has834834discordant pairs in M3\-occ’s favor and0against,p<10−6p<10^\{\-6\}\. Because the ground truth uses fine analytic sampling and M3\-occ uses a coarse DDA, the two implementations of visibility are independent, and the win is not an artifact of scoring a method against its own predicate\.

![Refer to caption](https://arxiv.org/html/2606.10299v1/x2.png)Figure 3:Pilot 2 \(controlled voxel sim\), occlusion over a wall slab with a doorway\. Left: accuracy on the849849occluded targets, text \(M0\) and the FoV cone \(M3a\) both0\.0000\.000, cone\-plus\-DDA\-line\-of\-sight \(M3\-occ\)0\.9820\.982\. Right: the false\-visible rate on occluded targets,1\.0001\.000for M0 and M3a \(every behind\-wall target is wrongly called visible\) versus0\.0180\.018for M3\-occ\. The live system’s*perception*is occlusion\-blind \(it cannot distinguish a behind\-wall target from a visible one\); the cheap DDA primitive recovers the visibility a content\-only index cannot compute\.
### 6\.3What the perception pilots show together

The perception pilots bracket the asymmetry\. On the trivial case, coordinate recall resolves collisions text cannot \(1\.0001\.000versus0\.5330\.533\)\. On the hard case, the live system’s geometry \(a FoV cone with no line\-of\-sight\) is exactly as blind as text \(0\.0000\.000on occluded targets, identical to M0\), and the cheap DDA primitive recovers it \(0\.9820\.982\)\. Both tests are decisive atp<10−6p<10^\{\-6\}with fully one\-sided discordant pairs\. They are small and controlled by design; their job is to confirm the asymmetry is real and measurable, and to size the confirmatory study\.

### 6\.4Tier\-A battery: the asymmetry holds across six geometry\-necessary types

To check that the asymmetry is not an artifact of one query type, we scaled the two pilots to the full Tier\-A taxonomy: six geometry\-necessary query types and one text\-sufficient control,400400trials each \(2,8002\{,\}800total\), on controlled voxel worlds with the same world\-geometry oracle\. Figure[4](https://arxiv.org/html/2606.10299#S6.F4)reports a geometry\-aware arm against a text/no\-geometry null\. On every geometry\-necessary type the geometry arm beats the null and the gap survives Holm\-Bonferroni correction over the six\-contrast family: presence1\.001\.00versus0\.500\.50, occlusion0\.990\.99versus0\.580\.58, relation1\.001\.00versus0\.260\.26, coverage/negation1\.001\.00versus0\.710\.71, near\-duplicate1\.001\.00versus0\.520\.52, and multi\-region1\.001\.00versus0\.510\.51\(all exact McNemarp<10−6p<10^\{\-6\}\)\. The text\-sufficient control inverts as it should: there the text null scores1\.0001\.000and the geometry arm0\.140\.14, so geometry does not masquerade as a general improvement \(the pre\-committed control\-inversion check F5 is negative, and F1 isolation\-fails and F2 null\-suffices are both empty\)\. We are explicit that these wins are partly by construction: the geometry arm computes the same geometric predicate the world\-geometry oracle uses to author ground truth, while the text null has no access to coordinates or occluders, so the gap on geometry\-necessary types is a structural fact \(text cannot compute occlusion, coverage, relation, or location\), not an estimated field effect size\. The battery shows the asymmetry holds across all six geometry\-necessary types AND that the control inverts \(F5 negative, geometry at chance on the text\-sufficient control\), demonstrating the instrument is not biased toward geometry; field effect sizes with live agents and blind raters are the confirmatory study, not this battery\.

![Refer to caption](https://arxiv.org/html/2606.10299v1/x3.png)Figure 4:Tier\-A battery \(400400trials per type\)\. Geometry\-aware recall versus a text/no\-geometry null across six geometry\-necessary query types and one text\-sufficient control \(T1\)\. Geometry wins every geometry\-necessary type \(all Holm\-corrected,p<10−6p<10^\{\-6\}\); on the text\-sufficient control the text null wins and geometry sits at chance, so the gain is specific to the geometry text cannot compute, not a general advantage\.
### 6\.5The memory\-method landscape cannot read geometry

The pilots so far compared geometry against a single text null\. To check that the asymmetry is not an artifact of a weak baseline, we ran the full text/RAG/agent\-memory method landscape with faithful cores, BM25 \(Okapi\), dense retrieval overbge\-small\-en\-v1\.5, hybrid reciprocal\-rank fusion, RAPTOR\-stylekk\-means cluster routing, Generative\-Agents scoring, and HippoRAG personalised PageRank, against a geometry\-aware arm, across three content domains \(village, library, workshop\) and the four geometry\-necessary query types,160160trials per cell\. Ground truth is computed from world geometry and is independent of every method \(non\-circular by construction\)\. Figure[5](https://arxiv.org/html/2606.10299#S6.F5)reports the pooled result\. On geometry\-necessary queries every method in the landscape sits at chance \(BM250\.4700\.470, dense0\.4610\.461, hybrid0\.4690\.469, RAPTOR0\.4730\.473, Generative\-Agents0\.4730\.473, HippoRAG0\.4590\.459\) while the geometry\-aware arm scores1\.0001\.000\. The floor is not a matter of retriever sophistication: lexical, dense, hybrid, hierarchical\-cluster, and graph memory all collapse to chance because the discriminating signal, occlusion and location and relation, is simply absent from the content they rank\. We are explicit that this gap is partly by construction: the discriminating signal \(coordinates, occluders\) is withheld from the content these methods rank, so a coordinate\-free retriever cannot exceed chance on a coordinate\-dependent query almost by definition\. The contribution here is a calibrated measurement of the floor any content\-only retriever hits on geometry\-necessary queries, not a competitive defeat of RAPTOR, GraphRAG, or HippoRAG on a task they were designed for\. That the same methods are competitive \(0\.570\.57to0\.770\.77\) on the text\-sufficient control shows the instrument is fair, not rigged\. The geometry arm drops to chance \(0\.2460\.246\) on the text\-sufficient control, confirming the gap is specific to geometry and not a general advantage\. The result holds in all three datasets\.

![Refer to caption](https://arxiv.org/html/2606.10299v1/x4.png)Figure 5:The full memory\-method landscape versus a geometry\-aware arm, pooled over three datasets and four geometry\-necessary query types \(160160trials per cell\)\. Every text/RAG/graph method is at chance on geometry\-necessary queries; only geometry\-aware recall succeeds\. On the text\-sufficient control the text methods are competitive and the geometry arm is at chance\. Ground truth is computed from geometry, independent of every method\.
### 6\.6L3: world\-bound memory changes action \(a flattening\-ablation pilot\)

The asymmetry above is a property of perception\. The claim that motivates the whole program is one level up: that binding a memory to its object changes what the agent*does*, not only what it retrieves\. We tested it with the decisive ablation directly\. Holding the world, the agent, and the task fixed, we varied only whether a memory was world\-bound \(surfaced because the action targets the object it is about\) or flattened to a text store \(surfaced by top\-33cosine over the flat corpus\), with each constraint phrased without the action’s keywords so that recall, not wording, decides\. A competent language agent \(Claude Sonnet\) then chose toproceedorrefusegiven only the task and what it recalled, across2424situated\-warning scenarios \(1616with a real object\-bound constraint,88benign controls\),4848decisions in total\. Figure[6](https://arxiv.org/html/2606.10299#S6.F6)reports action accuracy\. World\-bound memory acts correctly every time \(1\.0001\.000: it refuses all1616harmful edits and proceeds on all88safe ones, with no over\-refusal\)\. Flattening the identical facts drops action accuracy to0\.6250\.625\. The paired McNemar test over scenarios is decisive:99scenarios flip from correct under world\-bound to wrong under flattening and none flip the other way \(exactp=0\.0039p=0\.0039\)\. The failure is two\-sided and legible in the agents’ own reasons: flattening misses the object\-bound constraint \(77wrong proceeds, the agent writing that no recalled memory references the object and then breaking the world\), and it mis\-attributes a distractor constraint to the wrong place \(22controls wrongly refused, and one proceed that confabulated a different object’s note\)\. World\-bound recall is clean because it is keyed by the object the action targets\. By the litmus that defines the level, flattening the same memory to text loses37\.537\.5points of action quality, so the memory is actionably spatial, not merely stored spatially\. This is a pilot:2424scenarios, one model, a single\-turn decision rather than a full perceive–plan–act loop, an idealized world\-bound recall, and a top\-33text null\. A load\-bearing confound remains: the action gain may reflect*object\-keyed*retrieval, the memory is bound to the exact object the action targets, rather than spatial geometry per se\. Section[7](https://arxiv.org/html/2606.10299#S7)resolves this directly: an object\-keyed non\-geometric arm ties world\-bound \(E1,1\.0001\.000versus1\.0001\.000, McNemarp=1\.0p=1\.0\), so the situated\-action gain is*object binding*, and a large\-store depth sweep \(E2\) shows it is binding, not retrieval depth\. We therefore narrow the L3 claim to world/object\-bound memory; geometry\-specific value lives in the perception layer \(occlusion, near\-duplicate, multi\-region\), where object\-keyed text is itself at chance because the answer is in no stored note\. The confirmatory study runs the decision inside the live action loop and varies the model further\.

![Refer to caption](https://arxiv.org/html/2606.10299v1/x5.png)Figure 6:L3 flattening ablation \(2424situated\-warning scenarios,4848decisions\)\. Same world, agent, and facts; only the memory binding changes\. World\-bound memory acts correctly everywhere \(1\.0001\.000\); the identical facts flattened to text act correctly0\.6250\.625of the time, failing both by missing the real constraint and by mis\-attributing a distractor\. Paired McNemarp=0\.0039p=0\.0039\.
### 6\.7Beyond one task: a four\-task flattening\-ablation battery

To test whether the action effect is specific to warnings, we ran the same ablation on three further situated tasks \(the warning task above is the fourth\),1818scenarios per task:*return\-and\-resume*\(continue unfinished work versus start over\),*spatial causal debugging*\(fix an upstream root cause recorded in memory versus treat the visible symptom\), and*multi\-agent handoff*\(defer to another builder’s recorded ownership versus proceed\)\. The effect is real but bounded\. Return\-and\-resume replicates the warning result: world\-bound action accuracy1\.0001\.000versus flattened0\.6110\.611, paired McNemarp=0\.016p=0\.016, surviving Holm correction over the three\-task family; the flattened agent, missing the prior\-progress note, concludes there is nothing to resume and discards the work\. Pooled across all decisions the gap is significant \(0\.9630\.963versus0\.8150\.815,p=0\.039p=0\.039\)\. But two tasks are honest nulls: multi\-agent handoff shows no gap \(1\.0001\.000versus1\.0001\.000\), because the ownership note shares vocabulary with the request so flat retrieval surfaces it, and deferring to the owner is a safe default even without it; and spatial causal debugging is not significant \(0\.8890\.889versus0\.8330\.833,p=1\.0p=1\.0\), because the causal note does not reliably change the action and even world\-bound recall sometimes treats the symptom\. The honest reading: world\-binding improves action most when the decision hinges on a memory the flat retriever misses*and*that memory is phrased far from the eventual action query; it adds nothing when the safe default already aligns or when the recalled fact does not determine the act\. Two of four situated tasks replicate the L3 effect; two do not\. The battery inherits every limitation of the L3 pilot it extends: a single\-turn proceed/refuse decision rather than a live perceive\-plan\-act loop, an idealized world\-bound recall, an object\-keyed \(not geometry\-specific\) mechanism per E1, and automated parsing of the agent’s decision; the Claude Haiku replication \(Limitations\) retires only the single\-model concern\. The honest scope is that object\-binding helps action when the deciding memory is one flat retrieval misses and the safe default does not already align, replicated across two models and two of four tasks\.

### 6\.8The ranker\-blend dilution, measured \(a pre\-registered recall experiment\)

The index\-versus\-ranker distinction \(formalized in Section[3](https://arxiv.org/html/2606.10299#S3)\) is, until here, an argument\. This experiment measures it\. We test the exact claim that folding spatial proximity into a linear blend score beside recency and importance can dilute the spatial signal, by pre\-registering a recall test on relay\-captured, scripted agent memory and freezing the hypothesis and the falsification to git*before any arm was scored*\(commitfd098df6b; corpus SHA\-256cc82a1f0\.\.\.\)\.

#### Corpus and regime\.

The corpus is120120relay\-captured, scripted agent memories \(agent\-path\-captured in a controlled regime\),88activities performed at1515distinct locations each, captured through the liveappend\_memorypath with realbge\-small\-en\-v1\.5embeddings \(384384\-dim\) and subject\-anchored on a16×16×416\\times 16\\times 4m grid \(every pair at least1\.51\.5m apart, minimum observed1\.7321\.732m\)\. The text of each instance is a different natural sentence with genuine within\-activity semantic overlap \(never byte\-identical\), and the location token is stripped from every text, so only*location*can disambiguate which of the1515instances of an activity a query means\. To avoid rigging the test for or against spatial,occurred\_atis spread over4545simulated days andimportancevaries in\[0\.11,0\.95\]\[0\.11,0\.95\], so the recency and importance axes genuinely vary; the recall center is the target subject anchor plus an offset \(11m,33m,66m, far, or random\), never distance0\. This is the honest middle: not hash\-scattered \(which would rig against spatial\) and not placed on the recall center \(which would rig for it\)\. The primary metric is Hit@5: a trial scores11if its single ground\-truth memory appears anywhere in the top five results, and0otherwise\. We report Hit@5 rather than precision@5; with one relevant item per query the two differ by a constant factor and Hit@5 is the honest name\.

#### Pre\-registered falsification\.

The shipped\-blend claim is falsified if, at the near \(11–33m\) standpoint, the shipped hybrid blend \(hybrid\_full\) does not beat the pure\-vector baseline \(vector\_only\) on Hit@5 with paired Wilcoxon signed\-rankp<0\.05p<0\.05*and*a bootstrap95%95\\%CI on the mean delta that excludes0; or if the near\-to\-far gradient is flat; or if the win exists only when the query names the place\. The pre\-registered comparison is preciselyhybrid\_fullversusvector\_only, the shipped linear blend against a position\-blind baseline\.

Table 4:Wave E pre\-registered recall experiment, Hit@5 by arm and recall offset \(n=120n\{=\}120targets per cell\)\.vector\_onlyis flat at0\.3330\.333\(position\-blind: with no spatial term the recall center cannot change a pure\-cosine ranking\)\. The shipped linear blendhybrid\_full\(old weights0\.350\.35relevance/0\.30/0\.30spatial/0\.15/0\.15recency/0\.15/0\.15importance/0\.05/0\.05staleness\) sits at or below the pure\-vector baseline at every near offset\. Isolating the spatial signal \(hybrid\_spatial,0\.50\.5relevance\+0\.5\+0\.5spatial\) wins decisively, and re\-tuning the shipped blend to spatial\-dominant \(hybrid\_fullretuned,0\.40/0\.45/0\.08/0\.05/0\.020\.40/0\.45/0\.08/0\.05/0\.02\) recovers most of that win\. All arms show a clean near\-to\-far gradient \(the position\-blind baseline cannot\)\. At far, the spatial signal correctly collapses, the cue is uninformative when the standpoint is near no instance\.11m33m66mfar00\.20\.20\.40\.40\.60\.60\.80\.811Recall offset from targetHit@5vector\_onlyhybrid\_full\(old\)hybrid\_spatialhybrid\_full\(retuned\)Figure 7:Wave E pre\-registered recall experiment, Hit@5 versus recall offset for the four arms \(n=120n\{=\}120per cell; same data as Table[4](https://arxiv.org/html/2606.10299#S6.T4)\)\. The shipped blend \(hybrid\_full, old\) sits at or below the position\-blindvector\_onlybaseline at every near offset; geometry\-led weighting \(hybrid\_spatialand the retunedhybrid\_full\) wins decisively at near and collapses at far, where the cue is uninformative\.
#### Result: the shipped blend fails its own pre\-registered test\.

At the near standpoint \(11m and33m pooled,n=240n\{=\}240paired Hit@5 comparisons\), the shippedhybrid\_fullblend scores0\.2960\.296againstvector\_only’s0\.3330\.333: a mean delta of−0\.0375\-0\.0375\(negative\), Wilcoxonp=0\.306p=0\.306, Cohen’sd=−0\.066d=\-0\.066, and a bootstrap95%95\\%CI of\[−0\.108,\+0\.033\]\[\-0\.108,\+0\.033\]that includes0\. All three pre\-registered conditions to claim a win fail, so the verdict against the frozen criterion isnull/loss\. We report this plainly: the shipped linear blend does not beat a position\-blind baseline at the standpoint the whole design favors, and is if anything slightly worse \(Table[4](https://arxiv.org/html/2606.10299#S6.T4), Figure[7](https://arxiv.org/html/2606.10299#S6.F7)\)\. Because each query identifies the activity but strips the location token, a position\-blind retriever can recover the correct activity class but cannot distinguish among its1515location instances; under Hit@5 the within\-activity location\-blind baseline is therefore5/15=0\.3335/15=0\.333\. Thevector\_onlyarm sits exactly at this rate across every offset: it recovers the semantic activity but provides no spatial disambiguation at all, which makes it a fair, well\-calibrated baseline rather than a degenerate one\.

#### Diagnosis: recency and importance out\-vote spatial\.

The null is a verdict on the weight mix, not on the spatial signal\. Isolating spatial against the same baseline,hybrid\_spatial\(0\.50\.5relevance\+0\.5\+0\.5spatial, no recency or importance\) versusvector\_onlyat near gives mean delta\+0\.500\+0\.500, Wilcoxonp=6\.4×10−28p=6\.4\\times 10^\{\-28\}, Cohen’sd=\+0\.998d=\+0\.998\(very large\), and bootstrap CI\[\+0\.438,\+0\.563\]\[\+0\.438,\+0\.563\]excluding0: a decisive win\. So the spatial signal is overwhelming\. What sinkshybrid\_fullis that in this disambiguation regime recency and importance are target\-irrelevant axes \(which of the1515instances is correct is fixed by location alone\), yet together they carry30%30\\%of the score and out\-vote the spatial term at30%30\\%, pulling recent\-or\-important\-but\-wrong instances into the top55and pushing the spatially correct instance out\. The shipped blend does not merely fail to help; it falls below the pure\-vector baseline because it injects target\-irrelevant ranking noise\.

#### Fix and control: re\-tuning to spatial\-dominant flips null to win\.

We re\-tunedhybrid\_fullto spatial\-dominant weights \(relevance0\.400\.40, spatial0\.450\.45, recency0\.080\.08, importance0\.050\.05, staleness0\.020\.02\) and re\-ran on the identical corpus design through the same RPC with explicit weights\. The retuned blend versusvector\_onlyat near gives mean delta\+0\.3208\+0\.3208, Wilcoxonp<10−15p<10^\{\-15\}\(z=−8\.16z=\-8\.16\), Cohen’sd=0\.619d=0\.619, and bootstrap CI\[\+0\.271,\+0\.392\]\[\+0\.271,\+0\.392\]excluding0: the pre\-registered criterion is now met, a win\. Critically, running the old mix as a control on the same corpus reproduces the original null \(mean delta−0\.0375\-0\.0375,p=0\.306p=0\.306, CI includes0\), so the weight change, not anything else, is the cause of the flip\. Both weight sets are now live \(the relay scoring constant and the RPC default\)\. We are honest about the status of the confirmation: the re\-tune was run on the*same regime*as the pre\-registration \(same corpus design, same SHA, the falsification criterion reused verbatim\), not as a fresh pre\-registration, so it confirms the diagnosis rather than constituting independent confirmatory evidence\.

#### What this measures\.

This provides quantitative evidence for the index\-versus\-ranker distinction\. The ranker blend, spatial as one term among recency and importance, drowns the spatial signal and ties or loses to a position\-blind baseline; making spatial dominant \(geometry\-led, spatial\-dominant ranking, where geometry leads the score rather than contributing a diluted term\) wins by half a point of Hit@5 at a near standpoint with a near\-unit effect size\. This is the geometry\-led\-ranking endpoint of the distinction, not a true index/gate \(which would cut the candidate set by geometry before ranking\), and that stronger form remains future work\. The honest pre\-registered null on the shipped blend is itself a methodological result: the field’s memory\-palace default, a linear blend of relevance, recency, importance, and a spatial proximity term, can be net\-negative relative to pure vector when the non\-spatial axes are target\-irrelevant, and a study that pre\-registers that exact shipped blend can faithfully report a loss\. The caveats are real and we keep them: a single embedding model \(bge\-small\-en\-v1\.5\), scripted deterministic grid placement rather than organic agent wandering, and a single query regime \(pure spatial disambiguation\)\. The re\-tune fixes this regime by re\-weighting alone; whether one fixed weight set serves proximity\-only, semantic\-only, and temporal queries at once is open, and query\-adaptive weighting is the correct long\-term direction \(Section[8](https://arxiv.org/html/2606.10299#S8)\)\.

### 6\.9Live pre\-registered occlusion confirmation \(E4\)

Pilot 2 ran in a controlled simulation;SPMEM\-OCC\-LIVE\-v1\(hypothesis, arms, ground\-truth method, and falsification git\-committed before any data\) runs the same test on the*live*relay recall path\. On eight deterministic worlds the agent builds, through its own tools, a≥\\geq2\-cell\-thick occluder into the relay’s session overlay, then appends twelve behind\-wall and twelve open\-line\-of\-sight memories \(realbge\-smallembeddings\) anchored at the subject cell\. Ground truth is an*independent*0\.050\.05m analytic ray\-march over the world geometry, not the relay’s0\.40\.4m DDA, so the visibility predicate is never scored against its own implementation\. Two distinct facts must be kept apart\. First, memory recall correctly*returns*all9696behind\-wall memories in every arm: a location\-indexed store should surface what is stored at a place even when a wall stands between, so recall is occlusion\-blind as required and is never the quantity under test\. Second, the measured quantity is the*perception*judgment “is this target visible from the standpoint”, and the three arms differ only in the line\-of\-sight flag that answers it: a cosine null \(M0, which stores no geometry and so can never mark a target occluded, defaulting every retrieved target to visible\) and the live FoV\-cone visibility query \(M3a\) both*answer wrong*, calling100%100\\%of the9696behind\-wall targets visible, while the line\-of\-sight query \(M3\-occ\) answers right for exactly those and only those \(false\-visible1\.000→0\.0001\.000\\to 0\.000\), with no change on the9696open\-line\-of\-sight controls \(TOST\-equivalent in every world\)\. The pooled exact McNemar over the occluded subset isb=96b\{=\}96,c=0c\{=\}0,p=2\.5×10−29p\{=\}2\.5\\times 10^\{\-29\}, and M3\-occ agrees with the independent oracle on every target\. Building the live test first surfaced and fixed a defect the simulation could not: the relay’s query\-path line\-of\-sight predicate marched to the stored*observer*anchor \(or null for subject\-anchored memories\) rather than the*subject*, so the occlusion\-aware visibility query was a silent no\-op for exactly the subject\-anchored memories it should evaluate; the recall RPCs now return the coalesced subject\-preferred anchor, and a positive\-control preflight that aborts unless the predicate provably reports a behind\-wall subject not visible guards every run\. The occlusion\-blindness of perception, its fix, and the predicate’s correctness are therefore properties of the live system, established under pre\-registration\. The worlds are scripted \(one occluder family, jittered per world\); the full human\-authored multi\-world study with blind raters and a text\-plus\-coordinates LLM arm remains the confirmatory ceiling above this live pilot\.

## 7Robustness and confound checks

We pre\-registered \(SPMEM\-ZERO\-REAL\-PREREG\-v1, v3 round\) and ran five checks that attack the pilots’ weakest points; the fifth, the live occlusion confirmation \(E4\), is reported as a main result in Section[6\.9](https://arxiv.org/html/2606.10299#S6.SS9), leaving four here\. Two narrow our own claims, which we report as the data demands\.

#### E1: the situated\-action gain is object binding, not geometry\.

A reviewer worry is that the L3 situated\-warning effect reflects object\-keyed retrieval \(the memory is bound to the exact object the action names\) rather than spatial geometry\. We added an object\-keyed but non\-geometric arm \(exact object\-id match, no coordinates or visibility\) to the situated\-warning task\. Across2424scenarios the object\-keyed arm scores1\.0001\.000, identical to world\-bound \(1\.0001\.000\) and far above flattened text \(0\.5420\.542; E1 is a separate re\-run of the warning task, so its flattened score differs from the original L3 run’s0\.6250\.625\); the paired McNemar of object\-keyed versus world\-bound is a perfect tie \(b=c=0b\{=\}c\{=\}0,p=1\.0p\{=\}1\.0\), while both beat flattened \(p<10−3p<10^\{\-3\}\)\. The honest conclusion: the situated\-action gain is*object binding*, which an object\-keyed index achieves with no geometry\. We therefore narrow the L3 claim to world/object\-bound memory and locate geometry\-specific value where object\-keyed text is itself helpless, the occlusion, near\-duplicate, and multi\-region queries whose answer is in no stored note\.

#### E2: retrieval depth does not rescue dense similarity\.

Against “a deeper retriever would close the gap”, we swept the flattened arm over top\-k∈\{1,3,5,10,20\}k\\in\\\{1,3,5,10,20\\\}and rankers \{cosine, BM25, hybrid\} in a realistic216216\-memory store\. Dense cosine surfaces the object\-bound constraint in only0\.0630\.063of cases,*flat across all depths*\(the note is phrased unlike the query and is swamped\); BM25 recovers it by top\-20 \(0\.9380\.938\), but only because the action lexically names the object; hybrid is between \(0\.8130\.813at top\-20\)\. World\-bound and object\-keyed are1\.0001\.000at every depth by construction\. Depth does not close the gap for dense retrieval, and closes it for BM25 only under lexical name overlap; binding by key is the robust mechanism\.

#### E3: storing geometry, not the medium, is what matters\.

We tested whether the occlusion advantage is about text\-versus\-geometry or about computation\. Four arms decide visibility on200200trials: geometry\-blind text with no coordinates scores0\.5100\.510\(chance,100%100\\%false\-visible on occluded targets\); text that stores the precomputed label is trivially perfect \(1\.0001\.000\); the DDA scores0\.9850\.985; and an LLM handed the observer, target, and wall coordinates*as text*derives line\-of\-sight at0\.9900\.990, statistically indistinguishable from the DDA \(McNemarp=1\.0p=1\.0\) and far above blind text \(p<10−6p<10^\{\-6\}\)\. This refines the thesis: “text cannot fake occlusion” is too strong, a reasoner handed the coordinates computes it\. The irreducible requirement is that the memory*store the geometry*so a computation can run; the cheap option is a one\-line ray\-march, the expensive option is per\-query LLM reasoning, and similarity\-only retrieval over captions, storing no coordinates and running no computation, stays at chance\. We state the caveat that makes this honest: the LLM arm is handed the coordinates and occluder as text, so the experiment establishes that, given stored geometry, the medium does not matter, not that text is sufficient for spatial memory\. We did not test whether a model can infer occluder geometry it was never given \(it cannot, from captions alone, which is exactly the geometry\-blind null’s failure\)\. The claim is precisely: store the geometry, then either a one\-line ray\-march or per\-query reasoning recovers visibility; store only captions and both stay at chance\.

#### E5: the visibility predicate is an approximation with honest failure modes\.

Against “occlusion is trivially solved by a deterministic predicate”, we stressed the DDA on eight worlds against a fine0\.010\.01m analytic ground truth\. A coarse0\.50\.5m ray\-march is near\-exact on solid walls, diagonal walls, multi\-wall, elevated targets, windows, and partial occluders \(0\.990\.99to1\.001\.00\), but it under\-samples geometry thinner than its own step: it steps*over*sparse pillars8\.8%8\.8\\%of the time \(false\-visible\), and on a0\.20\.2m\-thin wall it misses the occluder on53\.8%53\.8\\%of behind\-wall targets \(accuracy0\.4630\.463, the battery’s worst failure, worse than a coin flip\)\. A0\.10\.1m march recovers both \(thin wall1\.001\.00, pillars0\.980\.98\)\. The step size is a real accuracy/cost knob, so the predicate is a genuine approximation, not a free oracle\.

## 8Discussion

The pilots support a single calibrated claim, the one the title names: spatial memory’s value for a language agent is what it must*store*, the world geometry a content\-only memory does not keep and similarity retrieval cannot recompute from captions\. The cleanest such geometry is visibility\. That occlusion requires geometry is not itself a finding, it is near\-tautological, and prior render\-as\-recall systems\[[8](https://arxiv.org/html/2606.10299#bib.bib8),[9](https://arxiv.org/html/2606.10299#bib.bib9)\]already exploit it\. Our finding is narrower: store the geometry \(not render it, per E3\) and keep the visibility predicate a separate verification step, not a recall\-ranker term \(Section[6\.8](https://arxiv.org/html/2606.10299#S6.SS8)\)\. Read through the index\-versus\-ranker distinction, the result is not that proximity should be a score\. It is that a memory must store the coordinates and occluders so the agent can*verify*a visibility predicate, a perception query, over what it recalls; occlusion is the predicate a content\-only index has no way to supply, because it neither stores the geometry nor runs the computation, and E3 shows that once the geometry is stored, either a one\-line ray\-march or an LLM reasoning step recovers it\. Our live system had a FoV cone but no line\-of\-sight query, so on behind\-wall targets its perception was no better than a content index\. The fix is cheap: one re\-pointed ray\-versus\-voxel DDA over the primitive the gaze ray already uses\.

This also clarifies the honest scope of what topic organization buys\. Sorting documents into topic regions is a cheap candidate\-narrowing trick: it can match flat retrieval at lower cost, but it never reads coordinates and it answers nothing a text index could not\. That is useful and it is not spatial memory\. Geometry’s irreducible contribution is elsewhere, in occlusion, visibility, viewpoint, and location, the queries where the answer depends on where things are and what can be seen from where\. We do not claim more than the pilots show\. We do not claim a best\-of\-N retrieval ranking, and we do not claim to have proven spatial memory “wins” beyond the measured asymmetry: text and the FoV cone fail on occlusion, the DDA recovers it, and coordinate recall resolves near\-duplicates\.

The pre\-registered recall experiment \(Section[6\.8](https://arxiv.org/html/2606.10299#S6.SS8)\) sharpens this into a second calibrated claim about*how*geometry must enter recall\. Spatial proximity is a strong, large\-effect, gradient\-bearing recall signal: isolated, it lifts Hit@5 by half a point at a near standpoint with a near\-unit effect size, and the near\-to\-far gradient is clean and monotone\. But it carries that value only when geometry*leads*\. The shipped memory\-palace default, a single linear blend that treats spatial proximity as one ranker term beside relevance, recency, and importance, dilutes the signal: when the non\-spatial axes are target\-irrelevant they out\-vote the spatial term, and the blend ties or even falls below a position\-blind pure\-vector baseline\. We pre\-registered that exact shipped blend and reported its loss faithfully rather than swapping in the spatial\-dominant arm that would have manufactured a win\. The honest reading is the index\-versus\-ranker distinction made quantitative: spatial value is realized when geometry leads recall, here as a spatial\-dominant \(geometry\-led\) ranking, not as a small diluted score term, and a naive linear blend can be net\-negative\. We are precise about the form: the demonstrated fix is a geometry\-led ranking, still a ranker; a true index/gate that cuts the candidate set by geometry before ranking is a stronger version we did not run and leave to future work\. That a widely\-assumed default can hurt is itself a finding the field should weigh\. The same\-corpus re\-tune to spatial\-dominant weights restores the win and the old\-mix control reproduces the null, isolating the weight as the cause, but it fixes only this one disambiguation regime; the right long\-term mechanism is almost certainly query\-adaptive weighting, or an index/gate, letting geometry lead when the query is spatial and recede when it is not, which we flag as future work rather than claim\.

The perception pilots establish only the perception layer, the prerequisite that geometry carries information text cannot\. The larger target is*actionable world memory*: memory bound to an object, place, affordance, episode history, social provenance, and the current task, used to change what the agent*does*on returning to a place, not merely what it retrieves\. The decisive test of that claim is a flattening ablation, holding world, agent, trajectory, and task fixed while varying only whether memory is world\-bound or flattened to text, and scoring action quality on situated tasks \(return\-and\-resume, situated warning, spatial causal debugging, multi\-agent handoff\); if flattening loses nothing, it is not spatial memory\. A first behavioral pilot of this ablation is reported in Section[6\.6](https://arxiv.org/html/2606.10299#S6.SS6): world\-bound memory acts correctly everywhere while the identical facts flattened to text fall to0\.6250\.625\(p=0\.0039p=0\.0039\), a first piece of action\-level evidence that the perception result is a substrate, not the endpoint\. The full behavioral battery remains part of the pre\-registered study\.

#### Research outlook\.

The perception\-level substrate measured here points to one open research question: whether a shared, editable spatial memory in which each artifact records where it is, why it is there, and what it connects to is more inspectable to humans than flat list, feed, or search retrieval when agent\-generated artifacts grow large\. The present paper tests only the lowest\-level requirement, that the system reads geometry text retrieval cannot compute; connecting that substrate to human\-facing organization of agent\-made artifacts, and measuring whether it improves human oversight, is future work\.

## 9Limitations

These are pilots, and we hold the claims to pilot evidence\. Both are small \(150 trials and one occlusion world\) and the occlusion pilot is a controlled JavaScript voxel simulation that mirrors Zero’sisSolidAtgeometry; its purpose is to isolate the visibility asymmetry cleanly, not to estimate field\-scale effect sizes\. The asymmetry is not only simulated, however:SPMEM\-OCC\-LIVE\-v1\(Section[6\.9](https://arxiv.org/html/2606.10299#S6.SS9)\) runs the line\-of\-sight visibility query on the live relay under a git\-committed pre\-registration across eight worlds, where it correctly answers that all9696behind\-wall targets are not visible \(recall still returns those memories\) whereas the cosine null and the FoV cone wrongly call them visible \(false\-visible0\.0000\.000, pooled exact McNemarp=2\.5×10−29p\{=\}2\.5\{\\times\}10^\{\-29\}\) with no effect on the9696open\-line\-of\-sight controls, and where building the live test first surfaced and fixed a relay bug that had marched the predicate to the wrong anchor\. Those worlds are scripted \(one occluder family, jittered\) and the separation is partly structural, since a content\-only index cannot compute occlusion by construction\. What remains is the human\-authored scale: the full confirmatory study \(12 worlds, roughly2,4502\{,\}450queries, blind raters, a text\-plus\-coordinates LLM arm, exact tests with Holm\-Bonferroni correction and TOST non\-inferiority margins, and the five pre\-committed falsification conditions F1–F5\) is pre\-registered and not yet run\. We use a single embedding model\. Every result reported here is scored automatically: the perception pilots, Tier\-A battery, landscape, andSPMEM\-OCC\-LIVE\-v1against an independent world\-geometry ray\-march oracle, and the L3 action pilots by parsing the agent’s own proceed/refuse decision\. No human rater judged any reported outcome\. Automated oracle scoring is the right instrument for the structural visibility asymmetry these pilots isolate \(visibility is a deterministic geometric fact\), but situated\-action quality is ultimately a human judgment; the pre\-registered confirmatory study introduces blind human raters precisely to close this gap, and until it runs the action\-level claims rest on automated decision\-parsing\. Zero’s terrain is flat ground with a brush overlay, not sculpted Genesis voxel terrain, so all geometry comes from the brush, post, and region universe the relay actually samples, and we claim nothing beyond it\. The richer representations M2 \(object\-at\-location scene graph\) and M4 \(rendered ego\-view read by a vision\-language model\) need infrastructure the live system lacks and are Tier\-B future work\. The L3 flattening\-ablation result \(Section[6\.6](https://arxiv.org/html/2606.10299#S6.SS6)\) is likewise a pilot:2424scenarios, a single model, a single\-turn decision rather than a live perceive–plan–act loop, an idealized world\-bound recall, and a top\-33cosine text null\. The action effect is not specific to one model: the battery replicates under Claude Haiku \(return\-and\-resumep=0\.031p=0\.031, pooledp=0\.039p=0\.039, with the same task\-by\-task signature\)\. The confirmatory study runs the decision inside the action loop and sweeps retrieval depth and more models\. Human\-rater and multi\-user metrics in the confirmatory design remain dependent on recruiting raters and users\.

#### Dynamic world and geometry versioning\.

A deeper open problem the present work does not address: occlusion here is computed against the world’s*current*geometry\. A memory’s visibility can change after the memory is encoded if occluders are added or removed, so a line\-of\-sight predicate evaluated at recall time can disagree with what was true when the memory was formed\. A complete answer to “what spatial memory must store” must therefore version the geometry, store an encoding\-time geometry snapshot, or store stable object identity so the predicate can be re\-evaluated against the right world state\. We do not address this here; the schema in Table[1](https://arxiv.org/html/2606.10299#S3.T1)and the pilots assume a single static geometry, and geometry versioning is left to future work\.

## 10Conclusion

Anchoring a memory to a coordinate is only worth doing if the agent then reads geometry that text cannot supply\. The cleanest such geometry is occlusion\. We identified that Zero’s live perception had no visibility query \(recall is correctly occlusion\-blind, surfacing the behind\-wall memory\), showed a cheap DDA primitive that supplies the missing visibility answer, and confirmed the asymmetry in two controlled pilots: coordinate recall resolves near\-duplicates that text cannot \(1\.0001\.000versus0\.5330\.533\), and on the visibility judgment for behind\-wall targets text and the FoV cone are equally blind \(0\.0000\.000\) while cone\-plus\-LOS recovers visibility \(0\.9820\.982\), both atp<10−6p<10^\{\-6\}\. We also demoted a prior topic\-partition study as not spatial memory\. Spatial memory for language agents should be built around the geometry a geometry\-blind text index cannot compute, and verification of visibility is the load\-bearing case\. A first behavioral pilot shows the payoff at the level that matters: binding the same facts to their object rather than flattening them to text lifts situated\-action accuracy from0\.6250\.625to1\.0001\.000\(p=0\.0039p=0\.0039\), and a four\-task battery shows the effect is real but task\-dependent \(two tasks replicate, two are honest nulls; pooledp=0\.039p=0\.039\)\.

A pre\-registered recall experiment then measures the index\-versus\-ranker distinction directly and tightens the prescription on*how*geometry must enter recall\. Spatial proximity is a strong, large\-effect, gradient\-bearing recall signal \(isolated:\+0\.500\+0\.500Hit@5 at near, Cohen’sd≈1\.0d\\approx 1\.0,p=6\.4×10−28p=6\.4\\times 10^\{\-28\}\), but only when geometry leads\. The shipped linear blend, which treats spatial as one term beside recency and importance, failed its own frozen pre\-registered test \(meanΔ\\DeltaHit@5−0\.0375\-0\.0375,p=0\.306p=0\.306, CI including0\) and sat below a position\-blind pure\-vector baseline at the near standpoint, an honest null we report rather than bury\. Re\-tuning to spatial\-dominant \(geometry\-led ranking\) weights flipped null to win \(\+0\.3208\+0\.3208,p<10−15p<10^\{\-15\}, CI excluding0\), with the old mix run as a same\-corpus control reproducing the null and isolating the weight as the cause\. The lesson is methodological as much as empirical: the field’s memory\-palace default, a naive linear blend of relevance, recency, importance, and a spatial proximity term, can be net\-negative relative to pure vector when the non\-spatial axes are target\-irrelevant; spatial earns its keep when geometry leads \(as an index/gate or a spatial\-dominant score\), not as a diluted score term\. Query\-adaptive weighting, and the stronger index/gate form we did not run here, are the right long\-term mechanism and are left as future work\.

#### Pre\-registration\.

The full confirmatory study is specified and frozen asSPMEM\-ZERO\-REAL\-PREREG\-v1before any confirmatory data collection\. It holds world, trajectory, observation stream, and brain fixed and varies only the memory representation acrossK=12K=12authored worlds with two trajectories each \(roughly2,4502\{,\}450queries over six geometry\-necessary query types plus a text\-sufficient control\)\. Ground truth is authored by an independent geometry kernel that has no read path to the memory tables, plus blind human raters\. The analysis is frozen \(exact McNemar and exact Wilcoxon/permutation tests, Holm\-Bonferroni family\-wise correction, TOST non\-inferiority with a declared margin, effect sizes and confidence intervals always reported\), and five falsification conditions \(isolation fails, null suffices, geometry unread, ground\-truth contamination, control inversion\) are pre\-committed so the study can fail honestly\. The two perception pilots reported here are its first runnable milestone: they demonstrate that the oracle reads world geometry with no memory access, that a spatial arm demonstrably dereferences coordinates, and that the geometry\-on/geometry\-off isolation delta is real and large on both the trivial and the hard case\. Beyond perception, the confirmatory study scales the action\-level result piloted here \(Section[6\.6](https://arxiv.org/html/2606.10299#S6.SS6)\) into a full behavioral battery \(return\-and\-resume, situated warning, spatial causal debugging, multi\-agent handoff\) under the flattening ablation, so the full claim is tested at the level that defines spatial memory: whether world\-bound memory changes what the agent does\.

#### Reproducibility\.

The pre\-registered recall experiment is reproducible from artifacts committed to the project repository: the harness lives atscripts/agents/wave\-e/\(corpus build, arm runner, and the spatial\-dominant re\-tune runner\) and the frozen pre\-registration atdocs/zero/research/agent\-memory/eval\-run\-e/wave\-e\-prereg/\(the falsificationpre\-registration\.json, the corpus freeze, the ground truth, and the per\-arm results\)\. The pre\-registration is a git\-committed artifact \(commitfd098df6b; corpus SHA\-256cc82a1f0\.\.\.\), written and committed before any arm was scored\. These artifacts \(the Wave E harness and frozen pre\-registration, theSPMEM\-OCC\-LIVE\-v1pre\-registration, live harness, and per\-world results, and the raw result files for every pilot reported here\) are included as ancillary files with this submission, so every number in the paper can be checked against its committed artifact without repository access\. Pilot 1 ran against the live relay through an embedding\-and\-recall path that was cleaned to zero residue after the run; Pilot 2 is a self\-contained deterministic voxel simulation\.

## References

- \[1\]J\. S\. Park, J\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein\.Generative Agents: Interactive Simulacra of Human Behavior\.*arXiv:2304\.03442*, 2023\.
- \[2\]C\. Packer, S\. Wooders, K\. Lin, V\. Fang, S\. G\. Patil, I\. Stoica, and J\. E\. Gonzalez\.MemGPT: Towards LLMs as Operating Systems\.*arXiv:2310\.08560*, 2023\.
- \[3\]W\. Xu, Z\. Liang, K\. Mei, H\. Gao, J\. Tan, and Y\. Zhang\.A\-MEM: Agentic Memory for LLM Agents\.*arXiv:2502\.12110*, 2025\.
- \[4\]P\. Chhikara, D\. Khant, S\. Aryan, T\. Singh, and D\. Yadav\.Mem0: Building Production\-Ready AI Agents with Scalable Long\-Term Memory\.*arXiv:2504\.19413*, 2025\.
- \[5\]P\. Rasmussen, P\. Paliychuk, T\. Beauvais, J\. Ryan, and D\. Chalef\.Zep: A Temporal Knowledge Graph Architecture for Agent Memory\.*arXiv:2501\.13956*, 2025\.
- \[6\]Y\. Yang et al\.3D\-Mem: 3D Scene Memory for Embodied Exploration and Reasoning\.*arXiv:2411\.17735*, 2024\. CVPR 2025\.
- \[7\]Z\. Cai, Y\. Du, C\. Wang, and Y\. Kong\.Vision to Geometry: 3D Spatial Memory for Sequential Embodied MLLM Reasoning and Exploration\.*arXiv:2512\.02458*, 2025\.
- \[8\]Y\. Lu, Y\. Du, D\. Liu, Y\. Zhou, C\. Wang, and Y\. Yin\.GSMem: 3D Gaussian Splatting as Persistent Spatial Memory for Zero\-Shot Embodied Exploration and Reasoning\.*arXiv:2603\.19137*, 2026\.
- \[9\]J\. Park and H\. Kang\.RenderMem: Rendering as Spatial Memory Retrieval\.*arXiv:2603\.14669*, 2026\.
- \[10\]Altera\.AL et al\.Project Sid: Many\-agent Simulations Toward AI Civilization\.*arXiv:2411\.00114*, 2024\.
- \[11\]P\. Sarthi, S\. Abdullah, A\. Tuli, S\. Khanna, A\. Goldie, and C\. D\. Manning\.RAPTOR: Recursive Abstractive Processing for Tree\-Organized Retrieval\.*ICLR*, 2024\. arXiv:2401\.18059\.
- \[12\]D\. Edge, H\. Trinh, N\. Cheng, J\. Bradley, A\. Chao, A\. Mody, S\. Truitt, D\. Metropolitansky, R\. O\. Ness, and J\. Larson\.From Local to Global: A Graph RAG Approach to Query\-Focused Summarization\.*arXiv:2404\.16130*, 2024\.
- \[13\]B\. J\. Gutiérrez, Y\. Shu, Y\. Gu, M\. Yasunaga, and Y\. Su\.HippoRAG: Neurobiologically Inspired Long\-Term Memory for Large Language Models\.*NeurIPS*, 2024\. arXiv:2405\.14831\.
- \[14\]S\. Robertson and H\. Zaragoza\.The Probabilistic Relevance Framework: BM25 and Beyond\.*Foundations and Trends in Information Retrieval*, 3\(4\):333–389, 2009\.
- \[15\]G\. V\. Cormack, C\. L\. A\. Clarke, and S\. Buettcher\.Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods\.*SIGIR*, 2009\.
- \[16\]NVIDIA\.Cosmos 3: Omnimodal World Models for Physical AI\.*arXiv:2606\.02800*, 2026\.
- \[17\]F\.\-F\. Li and the World Labs Team\.A Functional Taxonomy of World Models\.World Labs essay, 2026\.