WikiMem: Wiki-Based Memory Systems

Updated 6 July 2026

WikiMem is a framework of wiki-based memory systems that transform Wikipedia’s structures into queryable databases and semantic networks for various applications.
It integrates relational schemas, graph algorithms, and governance signals to enable reliable information retrieval, auditing, and LLM-based privacy assessments.
WikiMem also supports low-latency LLM inference through persistent KV caches and refined taxonomic backbones to enhance scalability and factual preservation.

Searching arXiv for papers and exact topic coverage. arXiv search: query = "WikiMem" WikiMem denotes a cluster of wiki-centered memory constructions that appear across several research lines rather than a single canonical artifact. In the cited literature, it refers variously to a relational, query-oriented storage of Wikipedia; a dynamic semantic network built from Wikipedia pages, links, and external URLs; a contributor- and governance-aware view of information reliability in collaborative encyclopedias; a dataset and metric for auditing person-specific factual memorization in LLMs; and a compiled wiki artifact served through persistent KV-cache inference in LLM Wiki systems (Katz et al., 2015, Fuehres et al., 2012, 0805.4722, Staufer, 15 Jul 2025, Huerta, 8 May 2026). The common thread is the treatment of wiki-derived structure—not only article text, but also links, redirects, categories, paragraph boundaries, contributor roles, edit histories, or taxonomic relations—as an operational memory substrate.

1. Scope and major research usages

The term is used across technically distinct but conceptually related settings. Some works treat Wikipedia as a structured external memory for IR, recommendation, sentiment analysis, and search; others treat wiki-derived signals as indicators of reliability, taxonomic coherence, or model memorization. This suggests that WikiMem is best understood as an umbrella for systems that compile, organize, or audit wiki-based knowledge rather than as a single standardized platform.

Research line	Core object	Paper
Queryable Wikipedia memory	Relational schema over pages, redirects, categories, links, paragraphs	(Katz et al., 2015)
Semantic and temporal knowledge mapping	Query-specific or topic-specific graphs over Wikipedia and Web URLs	(Fuehres et al., 2012, Puspitaningrum et al., 2017)
Reliability and governance	Contributor typology, arbitration, and article banners as trust signals	(0805.4722)
LLM privacy auditing	Dataset of natural-language canaries plus calibrated memorization metric	(Staufer, 15 Jul 2025)
LLM wiki-memory serving	Compiled wiki artifact loaded once and reused via KV cache	(Huerta, 8 May 2026)
Taxonomic backbone refinement	Cleaned Wikidata hierarchy for downstream memory-like systems	(Peng et al., 2024)

Historically, the line from collaborative-web reliability studies to structured Wikipedia databases and graph-based search systems foregrounded Wikipedia as a socially produced but computationally exploitable knowledge source. More recent work shifts the same intuition into LLM infrastructure: either by using wiki-like artifacts as persistent context for inference or by using wiki-derived facts to probe what models memorize about identifiable people (0805.4722, Katz et al., 2015, Huerta, 8 May 2026, Staufer, 15 Jul 2025).

2. Wikipedia as a structured and queryable memory

A foundational technical instantiation of WikiMem is the relational schema presented in "Enabling Complex Wikipedia Queries" (Katz et al., 2015). Its objective is to turn Wikipedia from raw dumps into a directly queryable resource for query-intensive applications. The schema stores page titles and types, redirects, categories, links with anchor text and character positions, and paragraph boundaries with hierarchy levels. It is reported as a relational database of about 50GB including indices, designed to fit on a stand-alone server or even a laptop rather than requiring a big-data stack.

The central table, Tbl_Wiki_Page, stores Page_id, Page_name, File_path, Page_type, and Stemmed_name. Page roles are encoded as 1 for entity pages, 2 for redirect pages, and 3 for category pages. Additional tables capture aliasing through Tbl_Wiki_Page_Redirect, many-to-many category membership through Tbl_Wiki_Page_Categories, intra-Wikipedia edges through Tbl_Wiki_Page_Links, and paragraph segmentation through Tbl_Wiki_Page_Paragraphs. A key design choice is the shared character-offset coordinate system between link positions and paragraph boundaries, which makes structure-sensitive SQL possible.

The report formalizes the schema as a relational universe with page entities

$P = \{ (id, name, type, stemmed\_name, file\_path) \},$

redirect relations

$R(p_r, p_t, t),$

category membership

$C(p, c),$

link relations

$L(p_i, p_j, a, pos),$

and paragraph relations

$H(p, k, start, end, level).$

This organization supports queries such as category-based retrieval, cross-group linking, redirect counting, and extraction of links from a specified paragraph by range-joining Pos_in_page against Paragraph_start_pos and Paragraph_end_pos (Katz et al., 2015).

In this form, WikiMem is a preprocessed and structurally enriched memory. It stores canonical identifiers, aliases, taxonomic attachments, surface forms, and positional evidence in a single schema. The report explicitly notes use in recommender systems, information retrieval, and sentiment analysis, indicating that the memory abstraction is not merely archival but computationally active (Katz et al., 2015).

3. Graph-based semantic memory and exploratory retrieval

A second major usage treats WikiMem as a graph rather than a database. "Galaxysearch - Discovering the Knowledge of Many by Using Wikipedia as a Meta-Searchindex" constructs a semantic network from Wikipedia articles, backlinks, out-links, outbound Web URLs, edit histories, and WikiProject quality and importance ratings (Fuehres et al., 2012). For a query term, seed articles are obtained from Wikipedia search, expanded through neighbors, and assembled into a graph with Wikipedia nodes and Web nodes. Ranking then depends on a combination of bidirectional link structure, quality, importance, actuality, and graph centrality.

The bidirectional-link heuristic is central. Instead of using raw indegree, which the authors found to be distorted by globally popular pages, GalaxySearch retains an article–article edge only when $A \to B$ and $B \to A$ . URLs inherit scores from the citing articles, for example

$\text{URLQuality}(u) = \sum_{A \in \Gamma(u)} Q(A), \quad \text{URLImportance}(u) = \sum_{A \in \Gamma(u)} I(A), \quad \text{URLActuality}(u) = \sum_{A \in \Gamma(u)} E(A),$

where $Q(A)$ , $I(A)$ , and $R(p_r, p_t, t),$ 0 are article quality, importance, and recent-edit counts. The system exposes three modes: WikiSearch, a ranked list of Wikipedia pages and Web URLs; WikiMap, a temporal animation of the topic graph; and WikiPulse, a recency-oriented mode built around section-level change analysis (Fuehres et al., 2012).

"Wiki-MetaSemantik: A Wikipedia-derived Query Expansion Approach based on Network Properties" pushes the graph-memory idea into online query expansion (Puspitaningrum et al., 2017). It builds, on the fly, a query-specific ontology graph from Wikipedia article titles, categories, “See also” links, and in-text hyperlinks, following links up to 3 hops. Candidate root concepts are obtained by searching "query wikipedia" and retaining Wikipedia URLs. The system then selects the “best concept” by graph degree, computes degree centrality, closeness centrality, and PageRank over the chosen graph, intersects high-ranking nodes across the three lists, and combines them with Borda Count to produce expansion terms.

The reported outcome is both effective and computationally light. Compared with a Persian Wikipedia ontology QE baseline, runtime ratios range between 1:70 and 1:281 for $R(p_r, p_t, t),$ 1, and the system is therefore described as having online potential. It also shows better performance in a meta-search engine setup than in an individual search engine setup, with closeness receiving the strongest empirical weight among the three network properties (Puspitaningrum et al., 2017). Taken together, GalaxySearch and Wiki-MetaSemantik instantiate WikiMem as a dynamic local semantic field whose graph topology determines salience, relevance, and temporal change.

Another line of work treats wiki memory as inseparable from governance. "La fiabilité des informations sur le web" examines the French Wikipedia as a collaborative knowledge system in which reliability is shaped by contributor behavior, arbitration, and community labeling rather than by direct expert signature (0805.4722). The paper does not attempt to verify factual correctness article by article. Instead, it studies structural reliability: where disputes occur, who participates in them, and how publication is controlled in conflictual pages.

The dataset combines a French Wikipedia snapshot saved on 2 April 2006 with two main sub-corpora: 1,000 articles that have or had a neutrality dispute banner, together with associated talk pages, and 80 arbitration cases up to April 2006. The snapshot contains more than 600,000 pages in total, about 370,000 article pages, and more than 40,000 article talk pages. Approximately 1,600 contributors appear in the non-neutral corpus, and these 1,600 contributors edited about 300,000 of the 370,000 articles, or about 81%, indicating that actors in neutrality disputes are central to overall content production (0805.4722).

The paper constructs a contributor typology along three axes: appearance before the Arbitration Committee, total volume of interventions, and focus on articles versus talk pages. Arbitration protagonists are divided into Très habitués with 3–14 arbitrations, Habitués with 2 arbitrations, and Occasionnels with 1 arbitration. Contribution volume is split into Très gros contributeurs with about 12,000–40,000 edits, Gros contributeurs with 2,800–12,000 edits, Contributeurs moyens with 600–2,800 edits, and Petits contributeurs with 1–600 edits. Almost all conflictual contributors primarily edit article space rather than discussions: 100% of very large contributors, 96% of large contributors, 81% of medium contributors, and 70% of small contributors focus mainly on articles (0805.4722).

Several empirical regularities are emphasized. Very frequent Arbitration Committee participants are mostly large contributors, with none classified as small. By role, very large contributors are 57% complainants, 29% accused, and 14% both, whereas small contributors are only 15% complainants and 75% accused. Among the 20 top contributors to protected articles, 35% have appeared before the Arbitration Committee; among top contributors to quality articles, only 3 are involved in arbitrations and they appear only as complainants. Across protected, non-neutral, and non-pertinent sets, most top contributors have special status such as administrator, arbitrator, or “wikipompier” (0805.4722).

The paper argues that high-volume contributors, especially those with special status, tend to have internalized neutrality, encyclopedic pertinence, and wikilove, and thus play a role of quality and editorial control. Low-volume contributors are more often the targets of complaints. This suggests a governance-based notion of WikiMem in which reliability is indexed by contributor composition, arbitration history, and community banners such as neutrality dispute, pertinence dispute, protection, or quality. The paper does not define a numeric reliability score, but it explicitly proposes these structural features as exploitable indicators for information systems (0805.4722).

5. WikiMem as an LLM memorization audit for personal data

In "What Should LLMs Forget? Quantifying Personal Data in LLMs for Right-to-Be-Forgotten Requests," WikiMem is the name of a dataset and evaluation framework for measuring which factual associations about identifiable persons are memorized by LLMs (Staufer, 15 Jul 2025). It is built from Wikidata and contains 243 human-related properties, 5,650 natural-language canaries, and for each property 100 counterfactual human–value pairs used as alternatives in value ranking. The motivating problem is that machine unlearning methods assume a known forget set, whereas GDPR and the Right to Be Forgotten require identifying which person-specific facts a model actually stores and can reveal.

WikiMem starts from Wikidata triples $R(p_r, p_t, t),$ 2 where $R(p_r, p_t, t),$ 3 is a human subject, $R(p_r, p_t, t),$ 4 is a property, and $R(p_r, p_t, t),$ 5 is a value. Properties are filtered to human subjects (P31 = Q5), restricted to datatypes that can be verbalized in prompts (wikibase-item, string, quantity, time), and retained only if they appear with at least 100 distinct humans. The result is a broad property inventory spanning occupation, citizenship, demographics, family relations, affiliations, and legal status. Each triple is rendered as a natural-language canary by a template $R(p_r, p_t, t),$ 6, yielding baseline declarative forms, paraphrased variants, and contextualized canaries that prepend auxiliary true facts about the same person (Staufer, 15 Jul 2025).

The metric is model-agnostic and likelihood-based. For a canary sentence tokenized as $R(p_r, p_t, t),$ 7, the negative log-likelihood is

$R(p_r, p_t, t),$ 8

Candidate values are scored by a calibrated difference that removes generic-person and name-string biases: $R(p_r, p_t, t),$ 9 where $C(p, c),$ 0 is a generic subject placeholder such as “This person’s,” $C(p, c),$ 1 is a set of similar-looking synthetic name variants, and $C(p, c),$ 2 in the experiments. A subject–property association is treated as memorized when some ground-truth value attains rank 1 among the true and counterfactual values; a strict subject–property decision further requires rank-1 success across all 11 prompt variants (Staufer, 15 Jul 2025).

The framework also defines a memorization-strength statistic. If $C(p, c),$ 3 is a top-ranked ground-truth value and $C(p, c),$ 4 is the highest score among non-ground-truth candidates, then $C(p, c),$ 5. Letting $C(p, c),$ 6 and $C(p, c),$ 7 denote the mean and standard deviation of analogous margins across candidate values, the standardized strength is

$C(p, c),$ 8

This yields both binary and graded views of person-specific factual memory (Staufer, 15 Jul 2025).

The experimental study evaluates 200 individuals, split into 100 well-known and 100 lesser-known subjects, across 15 causal LMs ranging from 410M to 70B parameters. For the five most frequent properties—occupation, language(s), place of birth, sex or gender, and citizenship—each subject–property pair is tested with one base template plus ten paraphrases against 100 counterfactual values. Memorization rates are consistently higher for well-known subjects and for larger models. For example, LLaMA 3.1-8B shows $C(p, c),$ 9 and $L(p_i, p_j, a, pos),$ 0 for well-known subjects versus $L(p_i, p_j, a, pos),$ 1 and $L(p_i, p_j, a, pos),$ 2 for lesser-known subjects; Pythia-410M shows 10.22% versus 4.38%; and LLaMA 3.1-70B shows 38.33% versus 18.66% (Staufer, 15 Jul 2025). Instruction tuning generally increases memorization for most properties, while “Sex or Gender” is reported as an exception, dropping near zero in the instruct-tuned model. Somewhat counterintuitively, contextualized canaries lower memorization rates and strengths across all models (Staufer, 15 Jul 2025).

In this usage, WikiMem is a privacy-auditing memory probe. It does not store knowledge for downstream retrieval; instead, it measures whether an LLM already stores factual associations of the form $L(p_i, p_j, a, pos),$ 3. The paper positions this directly as a mechanism for constructing individual-level forget sets for machine unlearning and RTBF workflows (Staufer, 15 Jul 2025).

6. Compiled wiki-memory for LLM inference

A further transformation of the idea appears in "WiCER: Wiki-memory Compile, Evaluate, Refine Iterative Knowledge Compilation for LLM Wiki Systems" (Huerta, 8 May 2026). Here, wiki-memory is an architectural pattern derived from Karpathy’s “LLM Wiki” idea and Chan et al.’s Cache-Augmented Generation: compile a domain’s documents into a persistent textual wiki, load that wiki once into an LLM’s context, cache the KV states, and answer subsequent questions against the cached prefix without retrieval at query time. In the reported implementation, Llama 3.1 8B Instruct with a 96k context is served through llama.cpp with prompt caching and quantized KV caches.

The attraction of the pattern is low-latency, retrieval-free serving. On the curated Policygenius corpus, full-context KV-cache inference outperforms RAG in both quality and time to first token: mean score 4.38 versus 4.08 out of 5, with warm-cache median TTFT 0.857 s versus 6.277 s, or 7.3 times faster. On the same setup, full-context score-1 rate is 1.2% compared with 5.8% for RAG (Huerta, 8 May 2026). However, on RepLiQA topics with 80 raw documents per topic and 55k–95k tokens, the pattern degrades: full-context raw reaches mean 3.47 for the 15 topics that fit into 96k, while RAG reaches 3.64 across all 17 topics. The paper attributes this reversal to attention dilution or “lost in the middle” effects (Huerta, 8 May 2026).

Blind wiki compilation fails even more severely. When an LLM compiler compresses the 80-document topics into “light,” “moderate,” and “aggressive” wikis, actual token ratios overshoot the targets and answer quality collapses. Relative to raw full-context at 3.46, the compiled conditions score 2.32, 2.25, and 2.14, with catastrophic score-1 rates of 52.9%, 57.1%, and 60.3%, respectively (Huerta, 8 May 2026). This is the paper’s “compilation gap”: naive compilation discards critical facts even though it yields very compact contexts.

WiCER addresses the gap with an iterative Compile–Evaluate–Refine loop inspired by counterexample-guided abstraction refinement. One diagnostic QA pair is selected per source document. A blind wiki $L(p_i, p_j, a, pos),$ 4 is first compiled at target ratio $L(p_i, p_j, a, pos),$ 5. The system then evaluates $L(p_i, p_j, a, pos),$ 6 on the probes, identifies catastrophic failures with score $L(p_i, p_j, a, pos),$ 7, diagnoses the critical facts in the corresponding source documents, accumulates those facts into a preserve set $L(p_i, p_j, a, pos),$ 8, and recompiles: $L(p_i, p_j, a, pos),$ 9 The process stops when failures disappear or improvement falls below 10% (Huerta, 8 May 2026).

Quantitatively, one to two WiCER iterations recover most of the lost quality. Across seventeen topics, the best WiCER wiki reaches mean 3.18; across the 15 topics with raw full-context baselines, the paper summarizes the result as mean 3.24 versus 3.47 for raw full-context, recovering 80% of the quality lost by blind compilation. The catastrophic failure rate drops from 55.1% to 24.8%, a 55% relative reduction (Huerta, 8 May 2026). An ablation shows that targeted diagnosis, not simply pinning extra random passages, drives the gain: random pinning improves mean score by only +0.16, whereas WiCER’s diagnosed pinning improves it by +0.95 across all 17 topics (Huerta, 8 May 2026).

This usage defines WikiMem as a persistent inference-time memory shard. The wiki text and its warmed KV cache function as the memory object; WiCER is the maintenance algorithm that makes such memory usable at scale by forcing preservation of facts that blind summarization would otherwise lose (Huerta, 8 May 2026).

A supporting but distinct infrastructure appears in "Refining Wikidata Taxonomy using LLMs," which presents WiKC, an automatically cleaned version of the Wikidata taxonomy (Peng et al., 2024). Although WiKC is not itself called WikiMem, the paper explicitly frames raw Wikidata as a problematic backbone for memory-like applications because of ambiguity between instanceOf and subclassOf, incorrect taxonomic paths, cycles, transitive redundancy, sparsity, and weak documentation.

The raw taxonomy is large and noisy: about 4.1M classes, about 4.8M subclassOf links, depth 20, many cycles, and only about 4% of classes instantiated. The paper gives concrete examples such as city inheriting through spatial entity and geometric object to mathematical object, and the cycle axiom → first principle → principle → axiom (Peng et al., 2024). To repair this, the authors use the Wikidata truthy dump of 22 March 2024, classify nodes into classes versus instances with explicit meta-class rules, build an initial acyclic taxonomy by DFS from entity, and then refine edges with Mixtral-8x7B-Instruct-v0.1 using the relation labels subclassOf, superclassOf, equivalent, irrelevant, and None.

The resulting WiKC taxonomy has about 17k classes, about 20k links, depth 13, no cycles, no redundant edges, no classes without instances, and 100% of classes with labels and descriptions. Relative to Wikidata, top-level classes drop from 38 to 16 and the average number of paths from an instance to root drops from 37 to 2.9 (Peng et al., 2024). In an extrinsic entity-typing evaluation judged by the same Mixtral model, macro accuracy rises from 43% for Wikidata to 70% for WiKC, with particularly large gains at depth $H(p, k, start, end, level).$ 0: 75% versus 37% (Peng et al., 2024).

For WikiMem-style systems, WiKC functions as a cleaned conceptual index. The paper explicitly links it to semantic search, question answering, entity recognition, entity linking, summarization, and LLM-augmented reasoning. This suggests a two-layer architecture in which wiki-memory systems use Wikipedia or compiled wiki artifacts for textual access while relying on a refined Wikidata taxonomy for stable subsumption structure and type constraints (Peng et al., 2024).

Taken together, these research lines define WikiMem not as a monolithic system but as a broader technical paradigm for wiki-based memory. Its implementations range from relational storage and graph ranking to governance-aware trust estimation, privacy auditing of factual memorization, persistent KV-cache artifacts for low-latency LLM serving, and cleaned taxonomic backbones. What unifies them is the assumption that wiki ecosystems encode reusable memory signals in their structure, not merely in their prose.