---
title: 'PReSO: Personalized Recall & Spatial Optimization'
url: https://www.emergentmind.com/topics/personalized-recall-and-spatial-optimization-preso
type: topic
---

# PReSO: Personalized Recall & Spatial Optimization

Searching arXiv for the cited PReSO-related papers to ground the article.
Personalized Recall and Spatial Optimization (PReSO) denotes, in its explicit formulation, the preprocessing workflow introduced in TourPlanner for constructing a spatially-aware candidate point-of-interest (POI) set before downstream itinerary generation. In that formulation, PReSO is designed to prune candidate POIs while maintaining a high recall rate, extract explicit demands and infer implicit user preferences, and produce a spatially compact, information-rich candidate set for subsequent reasoning and refinement [2601.04698]. A broader research reading is also suggested by adjacent work on recommendation, multimodal personal memory, embodied agents, and on-device retrieval: in those settings, PReSO can be understood as the joint problem of user-conditioned recall and some form of spatial structuring, where the “spatial” axis may mean geographic clustering, location-aware retrieval, embodied world coordinates, or structured organization of embedding space rather than a single universal optimization formalism [2104.07404], [2509.18436], [2606.03374].

## 1. Conceptual scope and meanings of “personalized recall” and “spatial optimization”

PReSO is not a single algorithmic template. In TourPlanner, it is a concrete workflow for candidate construction in travel planning. In nearby literatures, however, the same two ideas decompose differently. “Personalized recall” may mean recovering candidate POIs from a city inventory given explicit and implicit user demands, retrieving a user’s own multimodal memories for question answering, deriving a recall embedding from a ranking embedding in a recommender, or reconstructing latent user preferences from cross-session tool-use history. “Spatial optimization” may mean geographic clustering and route ordering, location-aware ranking over geotagged memories, explicit geometric predicates such as line-of-sight and occlusion, or a more limited structuring of retrieval space through basis vectors or cluster-guided search [2601.04698], [2509.18436], [2606.10299].

A recurrent misconception is that all “spatial” mechanisms are geometric in the same sense. The literature is more heterogeneous. In Memory-QA, the spatial contribution is location-aware retrieval through a lexical score over location strings, not map-based reasoning or geodesic optimization. In embodied-agent memory, by contrast, geometry becomes decisive when a query depends on observer pose, target location, and intervening occupancy. In recommendation, a paper such as UniRec is only indirectly related to spatial optimization: it creates a structured recall space through basis embeddings and top-\(P\) basis selection, but it does not optimize for index locality, quantization robustness, or explicit geometric constraints [2509.18436], [2606.10299], [2104.07404].

This suggests that PReSO is best treated as a family resemblance concept. Its strongest common denominator is the coupling of recall quality with a structured organization of the search space under system constraints. What changes across domains is the object being recalled—POIs, memories, user preferences, embodied observations—and the meaning of the spatial axis.

## 2. The explicit PReSO workflow in travel planning

In TourPlanner, PReSO is the first stage of a three-stage pipeline: PReSO constructs candidate information, Competitive consensus Chain-of-Thought (CCoT) explores the feasible solution space, and constraint-gated reinforcement learning refines the itinerary. PReSO itself is a sequential three-step workflow consisting of **User Profile Construction**, **Multi-dimension POIs Recall**, and **Spatial Clustering and Integration**. Its stated purpose is to solve the challenge of “Pruning candidate POIs while maintaining a high recall rate” and to “construct spatially-aware candidate POIs’ set” that is “spatially compact and information-rich” [2601.04698].

The first step, user profile construction, starts from a natural-language travel request \(Q\) and augments it with city-specific statistical data such as transportation prices, hotel price statistics, and restaurant meal price statistics. An LLM then infers latent requirements including hotel cost category and meal cost range. The appendix prompt rules make this explicit through heuristic budget allocation. Let \(N = \text{travel days} - 1\). Then the prompt uses
\[
\text{Per-night hotel budget} = \frac{\text{Budget} \times 0.55}{N}
\]
and
\[
\text{Per-day meal budget} = \frac{\text{Budget} \times 0.35}{N}.
\]
These rules are not presented as a learned objective; they are part of the prompt-guided inference mechanism.

The second step, multi-dimension POIs recall, uses a three-branch retrieval design. The first branch performs embedding-based semantic recall, broadening coverage with extracted keywords and synonym expansion. The second branch performs canonical landmark recall over attractions rated **4A or above**, ranking them by popularity and user ratings. The third branch is LLM-supplemented recall, which adds attractions that fit user preferences but may be missed by the first two channels. The appendix gives concrete sizing rules:
\[
\text{semantic similarity recall number} = 3 \times \text{duration},
\qquad
\text{POI recall total number} = 9 \times \text{duration}.
\]

The third step, spatial clustering and integration, clusters recalled attractions by geographic coordinates using DBSCAN, computes cluster centroids, and uses those centroids as spatial anchors for nearby accommodations and restaurants. The appendix reports DBSCAN hyperparameters of minimum samples \(=4\), \(\epsilon = 1\), and minimum cluster number \(=\) duration. The output is not just a ranked attraction list but clustered urban data with cluster labels attached to attractions, restaurants, and accommodations.

Empirically, the paper isolates PReSO by comparing its candidate recall against TripTailor on the TripTailor sandbox. With GPT-4o, PReSO reaches **42.26% recall** versus **27.83% recall** for TripTailor, an absolute gain of **14.43%**. The paper interprets this as evidence that the hybrid, multi-dimensional retrieval mechanism is better at capturing relevant environmental data before downstream reasoning [2601.04698].

## 3. Personalized recall as a systems problem

Outside itinerary planning, the recall side of PReSO appears in several distinct technical forms. In news recommendation, UniRec unifies candidate recall and ranking by learning a ranking user embedding \(\mathbf{u}_{ra}\) from click history and deriving a recall embedding \(\mathbf{u}_{re}\) through attention over trainable basis user embeddings. The attention weights are
\[
\alpha_i = \frac{\exp(\mathbf{u}_{ra}\cdot \mathbf{w}_i)}{\sum_{j=1}^M \exp(\mathbf{u}_{ra}\cdot \mathbf{w}_j)},
\]
and the recall embedding is synthesized as
\[
\mathbf{u}_{re} = \sum_{i=1}^M \alpha_i \mathbf{v}_i.
\]
At inference, only the top \(P\) basis embeddings are kept. With \(M=20\) and \(P=5\), UniRec(top) reaches **R@100 = 1.516**, **R@200 = 2.531**, **R@500 = 5.142**, and **R@1000 = 8.485** on MIND, while also slightly improving ranking over NRMS. The paper’s relevance to PReSO lies in the separation between a discriminative ranking space and a recall space optimized for broader interest coverage [2104.07404].

In personal multimodal memory QA, Pensieve formalizes a memory repository
\[
M = \{M_1, M_2, ... , M_n\}, \qquad M_i=(I_i, C_i, T_i, L_i),
\]
and augments it into
\[
M_i^a=(I_i, C_i, T_i, L_i, X_i),
\]
where \(X_i\) contains OCR output, image description, and invocation completion. Retrieval fuses temporal, recency, location, and semantic signals:
\[
s_i = w_t R_t (M_i, q, T_q) + w_r R_r(M_i, T_q) + w_l R_l(M_i, q) + w_s R_s(M_i^a, q).
\]
The learned weights reported are \(w_t=0.08\), \(w_r=0.22\), \(w_l=0.16\), and \(w_s=0.53\). On MemoryQA-s, Pensieve with GPT-4o reaches **76.8** end-to-end QA accuracy versus **63.1** for RagVL, and retrieval reaches **Recall@5 = 95.5** with learned-weight fusion. This is a strong example of personalized recall over a user-owned memory store rather than over a public corpus [2509.18436].

Cross-session preference recall appears in PRefine, which treats personalized tool calling as latent preference reasoning rather than direct fact retrieval. Its memory is a single evolving preference hypothesis \(M_T\), updated through a generate–verify–refine loop that checks Evidence Support, Abstraction Quality, Actionability, and Temporal Consistency. On the MPT benchmark of **265** multi-session dialogues, PRefine improves context-free Preference Recall average F1 from **53.19** to **64.51**, Preference Induction from **43.00** to **52.01**, and Preference Transfer from **16.26** to **19.65**, while using on average **23.28 tokens per dialogue**, or **1.24%** of the full dialogue history [2604.17886].

Adaptive retrieval depth appears in RF-Mem, which distinguishes a fast Familiarity path from a deeper Recollection path. The familiarity probe computes top-\(K\) similarity scores \(s_i\), normalizes them with
\[
p_i = \frac{\exp(\lambda (s_i - \max_j s_j))}{\sum_{j=1}^K \exp(\lambda (s_j - \max_j s_j))},
\]
then uses entropy
\[
H(p) = - \sum_{i=1}^K p_i \log p_i
\]
and mean score \(\bar{s}\) to switch between retrieval modes. The Recollection path clusters candidate memories, applies \(\alpha\)-mix query updates in embedding space, and performs bounded beam expansion. On PersonaMem at 128K scale, RF-Mem reaches **0.5394** accuracy versus **0.5259** for dense retrieval and **0.3231** for Full Context, while remaining far cheaper than full-history prompting [2603.09250].

Together these systems suggest that personalized recall in PReSO is not exhausted by top-\(k\) similarity search. It may require query decomposition, recall-space transformation, multimodal augmentation, latent preference abstraction, or adaptive control over retrieval depth.

## 4. Spatial optimization across geographic, location-aware, and embodied settings

The most literal form of PReSO’s spatial side appears in itinerary planning. ITINERA first decomposes a request into structured subrequests, retrieves POIs with positive-minus-negative semantic scoring, clusters retrieved POIs geographically, and then orders them using cluster-aware spatial optimization. Inter-cluster order is produced by a TSP solver, while within-cluster order is solved as a constrained path-TSP. The supplementary material specifies simulated annealing for cluster ordering with \(T_{\text{init}}=5000\), \(T_{\min}=0\), and cooling rate \(\alpha = 0.99\). On Shanghai, ITINERA reports **Recall Rate = 30.7%**, **Average Margin = 84.1 m**, and **Overlaps = 0.40**, substantially improving route coherence over strong LLM baselines; removing spatial optimization raises Average Margin to **227.0** and Overlaps to **1.00** [2402.07204].

TourPlanner’s explicit PReSO workflow is geographically simpler but conceptually similar. It uses DBSCAN clustering over recalled attractions and selects nearby restaurants and hotels around cluster centroids. Unlike ITINERA, it does not expose a full route-ordering solver inside PReSO itself; that burden is deferred to downstream CCoT and reinforcement-learning stages. This suggests two variants of spatial optimization within the PReSO family: cluster-based candidate shaping before reasoning, and explicit route optimization after semantic retrieval [2601.04698].

A weaker notion of spatiality appears in personal memory QA. In Pensieve, each memory includes a location field \(L_i\), but the location score is lexical:
\[
R_l(M_i, q) = \mathrm{BM25}(L_i, q).
\]
The paper is explicit that it does not use geographic distance, learned geospatial embeddings, or map-based reasoning. Here, “spatial” means location-aware metadata fusion rather than geometry. This matters because it marks a boundary: a system may be highly relevant to personalized recall while only partially addressing the spatial side of PReSO [2509.18436].

Embodied-agent memory pushes the spatial component much further. eMEM stores observations, gists, episodes, and entities in a unified graph backed by SQLite, HNSW, and an R-tree, making memory jointly searchable by meaning, space, and time. The system exposes ten agent-facing tools, including `locate`, which resolves a concept to a centroid and spread radius, and `recall`, which chains concept-to-location resolution with spatial and cross-layer retrieval. On eMEM-Bench v1, eMEM scores **80.8 weighted mean over 988 probes**, with a flat retention curve at ceiling from **1 h to 1 yr** of simulated delay on room-unique items. The abstract further reports that a pure RAG baseline loses **30 pt** on context-dependent retrieval and **29 pt** on DRM lure rejection, isolating the contribution of multi-layer storage and consolidation [2606.03374].

The strongest geometric interpretation is given by the study of language-agent spatial memory under occlusion. That work argues that geometry should lead retrieval when the query regime is spatial and that recall must be separated from visibility. A geometry-led weighting yields mean \(\Delta\)Hit@5 \(= +0.3208\) with \(p < 10^{-15}\), while the shipped linear blend fails its frozen test with mean \(\Delta\)Hit@5 \(= -0.0375\), \(p = 0.306\). For visibility, the paper repurposes a DDA line-of-sight test and shows that, on **849 behind-wall targets**, text-only and live field-of-view cone both score **0.000**, whereas cone-plus-DDA reaches **0.982**. This is a useful corrective to loose uses of “spatial optimization”: some spatial queries cannot be solved by location metadata or semantic similarity because they require explicit geometry and occupancy [2606.10299].

## 5. Resource-bounded recall, compression, and bounded activation

A distinctive feature of PReSO-style systems is that recall quality is constrained by deployment budgets. The literature therefore treats storage, latency, energy, and prompt bandwidth as part of the problem rather than as afterthoughts.

| System | Main mechanism | Reported effect |
|---|---|---|
| RECALL / GR-T [2409.15342] | Coarse-to-fine on-device multimodal embeddings | **14.9×** throughput, **13.1×** average energy reduction, **<5%** relative accuracy loss |
| ScrapMem [2605.03804] | Scrapbook Pages, EM-Graph, Optical Forgetting | **51.0% Joint@10**, **Recall@10 = 70.3%**, memory usage reduced by **up to 93%** |
| Memento [2504.19772] | Physiological cue selection for route recall | route recall improved by **20–23%**, review time reduced by **46%**, **3.86 s** vs **15.35 s** runtime |
| CBEA+LCV [2605.16712] | Bounded evidence activation and commitment validation | zero failures within validator scope at **0.49–0.60** availability; raw/long-context baselines with same gate reach zero only at **0.003–0.092** |

RECALL addresses the on-device substrate problem for personal multimodal recall. It computes coarse-grained embeddings via early exit during continuous ingestion, uses a lightweight predictor to estimate exit depth, and refines only shortlisted candidates at query time. In the paper’s synthesis, this supports a personal recall service that remains privacy-preserving and feasible on mobile hardware, with about **5 KB/item**, about **29.3 MB/day** at **6000 images/day**, and about **10.4 GB/year** of storage [2409.15342].

ScrapMem addresses the same problem from a different angle. It first consolidates heterogeneous personal records into temporally grounded Scrapbook Pages,
\[
\mathcal{S}_t = \mathcal{C}(\{I_i, V_j, T_k\}_t),
\]
then applies Optical Forgetting,
\[
\mathcal{S}_t^{(\tau)} = \mathcal{D}_{\tau}(\mathcal{S}_t),
\]
so that older memories are progressively degraded by JPEG-quality and resolution schedules. The default Timed-Gentle setting is \(Q = 90/70/40\), \(S = 1.0/0.85/0.60\), and \(T = 180/730\) days for Recent, Mid-term, and Old memories. The resulting graph shrinks from \(\mathcal{V}_t\) to \(\mathcal{V}_t^{(\tau)} \subseteq \mathcal{V}_t\), turning storage compression into a retrieval-aware structural condensation [2605.03804].

Memento studies a different resource bottleneck: how to select personalized route cues without heavy computer vision processing. It uses EEG, GSR, and PPG, fuses them with CPD and Morlet-wavelet analysis, and selects route snapshots aligned with high-attention episodes. Relative to the AMNet memorability baseline, it reports **3.86 ± 0.09 sec** runtime versus **15.35 ± 0.16 sec**, while improving route recall and lowering PAAS and NASA TLX. This is relevant to PReSO because it shows that a personalized recall aid can optimize not only storage but also review burden and cognitive load [2504.19772].

At the prompt/runtime layer, “Recall Isn’t Enough” argues that bounded evidence activation matters more than raw memory exposure. CBEA scores an evidence subset \(Z\) under relevance, coverage, tail witnesses, consequence debt, and overpersonalization penalty:
\[
J_t(Z)= \lambda_r\,\mathrm{Rel}(Z,c_t)
+\lambda_c\,\mathrm{Cov}(Z,R_t)
+\lambda_w\,\mathrm{Tail}(Z)
+\lambda_d\,\mathrm{Debt}(Z,u_t)
-\lambda_o\,\mathrm{Over}(Z,c_t),
\]
subject to a total activation-cost budget. Its lexicographic validator then ranks structured commitments by hard-constraint violations, coverage failures, infeasible emission, and soft utility. The paper’s central lesson for PReSO is that a personalized system may need to optimize which evidence is activated under finite budget before it can safely commit to an answer [2605.16712].

## 6. Evaluation, failure modes, and open directions

A major shift in recent work is that PReSO-like systems are increasingly evaluated as state-maintenance systems rather than as pure retrievers. Memora makes this explicit for personalized agents operating over weeks to months of interaction. It evaluates remembering, reasoning, and recommending under additions, updates, and deletions, and introduces Forgetting-Aware Memory Accuracy:
\[
\text{FAMA} = \max\!\Big(0,\; \text{MPA} - \lambda \cdot (1 - \text{FAA})\Big),
\qquad
\lambda = \frac{N_{\text{forget}}}{N_{\text{presence}} + N_{\text{forget}}}.
\]
Here MPA measures memory presence accuracy and FAA measures forgetting absence accuracy. The benchmark shows that external memory agents improve direct remembering but still frequently reuse obsolete memories; quarterly settings include maxima of **309 prior sessions** and **94 mutations** before a query. For PReSO, this implies that personalized recall must be update-aware and forgetting-aware, not just high-recall in the retrieval sense [2604.20006].

Benchmark design is likewise becoming more diagnostic. MemoryQA contributes **9,357 recall questions** with single-memory and multi-memory settings, explicit time-only, location-only, and joint time-and-location constraints. MPT contributes **332** Preference Recall, **293** Preference Induction, and **472** Preference Transfer instances for cross-session tool calling. eMEM-Bench is organized around eight cognitive-psychology paradigms rather than surface tasks. This diversity suggests that no single metric captures PReSO. A system may excel at semantic recall yet fail on temporal mutation, location grounding, lure rejection, or geometry-dependent visibility [2509.18436], [2604.17886], [2606.03374].

Several limitations recur across the literature. First, many systems are strong on personalized recall but only partially spatial. Memory-QA uses location-aware retrieval without geospatial reasoning; RF-Mem explores semantic neighborhoods without a persistent spatial memory layout; UniRec creates a structured recall space but does not optimize for index locality or hardware-aware retrieval [2509.18436], [2603.09250], [2104.07404]. Second, strong spatial systems are often not yet personalized: eMEM is explicitly a spatio-temporal memory substrate rather than a user-modeling system. Third, high recall can still produce bad commitments: Memora shows stale-memory reuse, while CBEA+LCV shows that failures often occur when noisy evidence is turned into hard commitments rather than at retrieval time [2604.20006], [2605.16712].

This suggests three open directions for PReSO. One is **joint personalization and geometry**, where user-specific salience, routines, and preferences are integrated with explicit spatial predicates rather than added as a post hoc reranker. A second is **budget-aware memory organization**, where retrieval quality, prompt cost, storage cost, and latency are optimized together. A third is **state-correct recall**, where systems must remember what is currently true, suppress invalidated facts, and abstain or repair when feasible commitments cannot be formed. The literature already provides strong ingredients for each of these directions, but no single system yet closes the full loop from personalized recall to explicit spatial optimization across all domains.

Source: https://www.emergentmind.com/topics/personalized-recall-and-spatial-optimization-preso