The Immutable Past: Formalizing State Mutability and Conflict Resolution in Mutable RAG

Hamed Haddadpajouh Affiliation: Independent Researcher Amir AmiriTabat Affiliation: Independent Researcher

Abstract

Retrieval-Augmented Generation (RAG) serves as the primary memory architecture for long-horizon autonomous agents. However, treating shared memory as an append-only stream introduces Semantic Shadowing, a critical failure mode where conflicting historical observations accumulate and statistically dominate valid recent updates. In dynamic environments, this results in severe state divergence as agents retrieve and act upon obsolete facts. This paper formalizes the mechanics of State Mutability to prove that standard dense retrieval suffers from Asymptotic Recall Decay. Furthermore, we formally demonstrate a Majority Vote Trap, revealing that increasing the retrieval context window paradoxically degrades generation accuracy by diluting the attention mechanism under conditions of semantic equivalence. To resolve this, we introduce GC-Mem (Garbage Collection for Memory), a strict inference-time consistency protocol. Unlike heuristic time-decay mechanisms—which indiscriminately destroy valid long-term memory—GC-Mem relies purely on a temporal dominance operator ( $\Phi_{\mathcal{T}}$ ) paired with contradiction detection to surgically excise shadowed context. Evaluated across a rigorous, behaviorally inferred benchmark of 137,760 memory chunks and continuous accumulation sweeps, standard RAG and timestamp re-ranking baselines experience severe degradation. In contrast, GC-Mem empirically recovers $>90\%$ conflict resolution accuracy. We establish strict precision and recall deployment thresholds, ensuring state convergence where standard mutable RAG fundamentally fails.

1 Introduction

As autonomous systems graduate from isolated tasks to long-horizon workflows, the consistency of their memory architecture becomes paramount. Retrieval-Augmented Generation (RAG) is the consensus memory framework Lewis et al. (2020), yet it operates under the implicit assumption of a static knowledge base. In real-world agentic deployments, memory is an append-only stream where factual states mutate over time (e.g., user preferences shift, project goals update, infrastructure changes) Park et al. (2023).

We identify a foundational failure in the append-only paradigm: when an entity updates, the historical state is not deleted. Because the historical and updated states are semantically equivalent but logically contradictory, dense retrieval algorithms return the statistical majority of stale chunks. We formalize this failure mode as Semantic Shadowing.

This paper establishes the theoretical bounds of retrieval failure under state mutation. We prove three central theorems: (1) valid-state recall decays asymptotically to zero as a system’s interaction history grows; (2) generation fails even when valid chunks are successfully retrieved because attention mass becomes non-discriminating across semantically equivalent contradictions; and (3) multi-factor memory scoring fundamentally violates state convergence by protecting entrenched historical errors.

To address these vulnerabilities, we propose GC-Mem, a lightweight, inference-time garbage collection operator driven strictly by temporal dominance ( $\Phi_{\mathcal{T}}$ ). Unlike naive timestamp re-ranking—which we demonstrate destroys valid historical reasoning—GC-Mem surgically removes logical contradictions from the retrieved context prior to generation.

2 Related Work

Agentic Memory and Mutable RAG. The deployment of long-horizon autonomous agents necessitates robust memory architectures. Current frameworks like Voyager Wang et al. (2023) and Reflexion Shinn et al. (2023) utilize iterative self-reflection to update agent policies, but rely on localized episodic buffers. MemGPT Packer et al. (2023) introduced an OS-inspired memory hierarchy to manage context limits, while GraphRAG Edge et al. (2024) and HippoRAG Gutiérrez et al. (2024) map global semantic structures for multi-hop reasoning. However, these architectures primarily treat memory as a monotonically growing, static knowledge base. When subjected to state mutation, standard dense retrieval Lewis et al. (2020); Karpov et al. (2024) falters because heuristic time-decay algorithms indiscriminately erase valid historical invariants alongside stale facts. GC-Mem explicitly solves the mutability gap in RAG-driven agent architectures.

Knowledge Editing and Temporal Contradiction. Our work bridges mutable RAG with the literature on knowledge editing and temporal QA. Methodologies like ROME Meng et al. (2022) and MEMIT Meng et al. (2023) directly mutate the parametric weights of LLMs to inject updated facts. Concurrently, TempLAMA Dhingra et al. (2022) and StreamingQA Liska et al. (2022) explore QA over temporally evolving corpora. While parametric editing is computationally prohibitive for continuous, high-frequency multi-agent streams, GC-Mem acts as a non-parametric, inference-time filter. By utilizing Natural Language Inference (NLI) contradiction detection, GC-Mem dynamically resolves temporal conflicts in the retrieved context without requiring weight updates.

Long-Context Dilution and Attention Collapse. The Majority Vote Trap formalized in this paper builds upon the "Lost in the Middle" phenomenon Liu et al. (2023); Levy et al. (2024), which demonstrated that LLM attention degrades when relevant information is flanked by noise. We extend this by proving that in mutable memory, context dilution is not an accidental byproduct of long documents, but a mathematical asymptote. As history grows, the probability of retrieving a correct state approaches zero, and the generation attention mass becomes captive to the statistical majority of stale evidence Wang et al. (2025).

Latency and Self-Correction in LLMs. Recent advancements in agentic reasoning heavily leverage critic-guided reflection loops Zhang et al. (2025) to correct hallucinations and reasoning errors. However, relying on multi-turn generative reflection to prune memory contradictions induces severe latency overhead and token costs. GC-Mem operates as a computationally lightweight, feed-forward inference filter. By excising logical contradictions prior to the primary generation phase, GC-Mem circumvents the need for post-generation reflection loops, ensuring strict factual grounding within bounded operational latency.

3 Theoretical Framework

Let $\mathcal{M}_{t}=\{(c_{1},\tau_{1}),\dots,(c_{n},\tau_{n})\}$ denote the shared memory stream at time $t$ , where $c_{i}$ is a memory chunk and $\tau_{i}$ is its timestamp.

3.1 Problem Formulation

Definition 3.1 (Mutable Entity State & Multi-Transition Dynamics).

Consider an entity $E$ . The ground truth state $\sigma_{E}(t)$ is formally defined as a piecewise-constant function over time. Let $\mathcal{T}=\{t_{0},t_{1},\dots,t_{n}\}$ denote a strictly monotonically increasing sequence of transition timestamps ( $t_{0}<t_{1}<\dots<t_{n}$ ). For any interval $t\in[t_{i},t_{i+1})$ , the entity state is $S_{i}$ .

For a query issued at current time $t>t_{n}$ , the terminal state $S_{new}=S_{n}$ is the sole valid state. All preceding states $S_{old}\in\{S_{0},\dots,S_{n-1}\}$ where $S_{old}\perp S_{new}$ constitute logical contradictions. This formalizes chained state mutations (e.g., $A\rightarrow B\rightarrow C$ ). Because the temporal operator $\Phi_{\mathcal{T}}$ enforces $\tau_{new}>\tau_{old}$ pairwise across all retrieved chunks, it dynamically collapses chained contradictions to the terminal valid state $S_{n}$ .

Lemma 3.2 (Semantic Equivalence).

For a query $q$ targeting the state of $E$ , the embedding similarity of the old and new states is mathematically indistinguishable up to a marginal noise term $\epsilon$ :

|sim(q,S_{old})-sim(q,S_{new})|<\epsilon

(1)

Assumption 3.3 (Accumulation Hypothesis ( $N\gg M$ )).

Let $\eta(S)$ denote the count of chunks asserting state $S$ semantically similar to a specific attribute. For instance, if an agent interacts with a user 50 times regarding their diet, and the diet mutates once, $N=49$ and $M=1$ . $N$ scales per-attribute, not globally. We assume $\eta(S_{old})\gg\eta(S_{new})$ .

3.2 The Shadowing Failure

Theorem 3.4 (Asymptotic Recall Decay).

Assuming the Accumulation Hypothesis ( $N\gg M$ ), the expected number of valid chunks $X$ in the retrieved set $\mathcal{R}_{k}$ is bounded by:

\mathbb{E}[X]\approx k\frac{M}{N+M}

(2)

As attribute history $N\to\infty$ , $\mathbb{E}[X]\to 0$ . Because dense retrievers exhibit inherent semantic bias toward dominant phrasings, this hypergeometric formulation serves as a theoretical worst-case lower bound for recall failure (Proof in Appendix B.1).

Theorem 3.5 (The Majority Vote Trap).

Even if $\mathcal{R}_{k}$ contains valid evidence (e.g., $k-1$ stale chunks and 1 valid chunk), generation collapses. Under Semantic Equivalence (Lemma 3.2), the LLM attention mechanism becomes non-discriminating across the chunks Liu et al. (2023). The probability mass assigned to the stale state scales with its representation:

P(y=S_{old})\propto\frac{k-1}{k}

(3)

(Proof in Appendix B.2).

Refer to caption — Figure 1: Embedding similarity noise ( $\epsilon$ ) distribution illustrating semantic equivalence between old and new states.

4 The GC-Mem Protocol

To mathematically guarantee state convergence, we introduce GC-Mem. Operationally, the system functions as a linear inference stack from ingested interaction logs to validated generation. We detail the architectural mechanics below.

4.1 Memory Architecture and Candidate Retrieval

Each multi-turn session is segmented and embedded via a fixed sentence encoder, populating the primary medium-term memory (MTM) vector index. Optionally, the architecture supports a Short-Term Memory (STM) recency buffer, implemented as a FIFO queue mirroring the $N$ most recent chunks. When a user query is issued, the system executes multi-query expansion to mitigate single-embedding bias. The MTM returns the top- $k$ semantically similar chunks. If the STM buffer is enabled, it is unioned with the MTM hits and deduplicated, forming the frozen candidate set $\mathcal{R}_{k}$ . Crucially, $\mathcal{R}_{k}$ serves as the identical baseline context evaluated across all methodologies (Standard RAG, Timestamp Re-ranking, and GC-Mem) to ensure strict experimental parity.

4.2 Conflict Oracle and Temporal Dominance

Each item is simplified to a tuple $m_{i}=(c_{i},\tau_{i},r_{i})$ , where $\tau_{i}$ is the strict chronological timestamp and $r_{i}$ is retrieval relevance.

We define the conflict function $\kappa(c_{i},c_{j})\in\{0,1\}$ which evaluates whether two retrieved statements logically contradict relative to the user query. This step utilizes a pairwise Natural Language Inference (NLI) heuristic operating strictly within $\mathcal{R}_{k}$ , resulting in a bounded $O(|\mathcal{R}_{k}|^{2})$ operational complexity.

We define the temporal dominance operator $\Phi_{\mathcal{T}}$ over the retrieved set $\mathcal{R}_{k}$ :

\Phi_{\mathcal{T}}(\mathcal{R}_{k})=\{c_{i}\in\mathcal{R}_{k}\mid\nexists c_{j}\in\mathcal{R}_{k}:\kappa(c_{i},c_{j})=1\land\tau_{j}>\tau_{i}\}

(4)

Theorem 4.1 (State Recovery).

Applying $\Phi_{\mathcal{T}}$ guarantees that if at least one valid chunk $S_{new}$ is retrieved, all older contradictory chunks $S_{old}$ are excised. The context is purged of contradiction, breaking the Majority Vote Trap and forcing $P(y=S_{new})\to 1$ .

5 Experimental Methodology

5.1 Dataset Scale and Rationale

We constructed a Temporal Mutation Benchmark consisting of 2,016 experiment instances and 137,760 evaluated memory chunks across continuous parameter sweeps (6 controlled $N:M$ accumulation ratios from 1:1 to 100:1, and 4 model architectures).

We utilize a controlled synthetic environment because organic data cannot isolate the mathematical bounds of Semantic Shadowing. We evaluated external datasets (FactConsolidation and WikiContradict) and found they yielded 0% and 10% Conflict Resolution Accuracy (CR-Acc), respectively, due to the complete absence of controlled temporal contradiction ratios.

5.2 The Behavioral Pipeline

To ensure robust evaluation, we utilize a behavioral generation pipeline to construct the memory contexts. Preliminary templates exhibited up to 14% keyword leakage (where the state is explicitly named, e.g., "Status: employed"). The pipeline relies exclusively on behavior-inferrable state constraints (e.g., Old: "Complained about feeling winded after climbing a single flight of stairs." vs New: "Jogged three miles before breakfast this morning."). Regulated by a secondary GPT-4 quality rater, keyword leakage is suppressed to 0-3%, forcing the retriever to rely purely on complex semantic similarity. The medium dataset contains 48 instances (1,464 chunks), and the large contains 65 instances (1,950 chunks).

Figure 2: System architecture: the end-to-end RAG system used in this paper (this figure depicts the system architecture, not the dataset-generation pipeline). The diagram shows chunking, embedding, the optional STM (short-term memory) FIFO, and the three parallel answer paths (Standard RAG, Timestamp re-rank, and GC-Mem). All approaches receive the same retrieved candidate set so differences are attributable to the conflict oracle

\kappa

and the temporal dominance operator

\Phi_{\mathcal{T}}

6 Results

6.1 Timestamp Baselines vs. GC-Mem

A common heuristic is that simple time-decay or timestamp re-ranking resolves mutable memory. Table 1 empirically refutes this. Timestamp heuristics blindly penalize all historical chunks, discarding valid, non-contradicted long-term facts alongside stale ones.

Model	Dataset Split	Standard RAG	Timestamp Re-ranking	GC-Mem ( $\Phi_{\mathcal{T}}$ )
deepseek-v3	Large ( $n=74$ )	56.8%	20.3%	97.3%*
o4-mini	Medium ( $n=38$ )	28.9%	10.5%	92.1%*
gpt-4o	Medium ( $n=38$ )	68.4%	10.5%	10.5%

Table 1: Conflict Resolution Accuracy (CR-Acc). Timestamps alone cause severe degradation. (* indicates

p<0.01

significance).

Figure 3: Effect of STM (Short-Term Memory) on conflict resolution accuracy across models: comparison between Timestamp re-ranking and GC-Mem. STM supplies recency evidence; the plot shows GC-Mem retains a substantial advantage because it excises contradictory old chunks rather than re-ranking them.

6.2 Short-Term Memory (STM) Dynamics

To evaluate architectural interaction, we activated the Short-Term Memory (STM) recency buffer. As demonstrated in Table 2, enabling STM raises the baseline accuracy of standard RAG by artificially elevating the exposure of new-state evidence in the retrieval window. However, GC-Mem retains a substantially wider performance gap. Because simple recency heuristics merely surface the new state alongside the vast quantity of historical states, the context remains contradictory. GC-Mem explicitly resolves this by pruning the semantically shadowed stale chunks that the STM forces into the context window.

Dataset	STM Buffer	Standard RAG	GC-Mem ( $\Phi_{\mathcal{T}}$ )	$\Delta$ GC Lift
Medium	STM=off	24.4%	40.2%	+15.8p
Medium	STM=on	46.1%	68.1%	+22.0p
Large	STM=off	15.9%	30.4%	+14.4p
Large	STM=on	34.5%	62.9%	+28.4p

Table 2: Headline accuracy averaged over the benchmark models. STM amplifies the value of GC-Mem by supplying the conflict oracle with richer temporal context.

6.3 Detector Dependency and Deployment Threshold

GC-Mem operates as a cognitive offloading protocol; its success is strictly bounded by the underlying model’s reasoning capabilities. Table 3 demonstrates that while precision is universally high, recall is the decisive factor.

Model	Precision	Recall	F1 Score	CR-Acc $\Delta$	Protocol Efficacy
deepseek-v3	98.9%	95.2%	0.97	+69.0 pp	Optimal
o4-mini	98.1%	99.8%	0.99	+79.0 pp	Optimal
gpt-4o	99.7%	36.6%	0.54	+4.0 pp	Marginal
gpt-4.1-mini	100.0%	0.7%	0.01	-0.2 pp	Fails

Table 3: Detector Dependency. High recall (

>50\%

) is strictly required for GC-Mem efficacy.

6.4 Ablation: The Failure of Multi-Factor Dominance

To evaluate whether multi-factor memory scoring helps resolve contradictions, we test a multi-factor scoring variant (e.g., scoring chunks using $\alpha(\text{recency})+\beta(\text{relevance})+\gamma(\text{usage})+\delta(\text{stability})$ ). We rigorously ablated this variant on the o4-mini architecture.

Variant	CR-Acc (Medium)	CR-Acc (Large)	Status
Eq. 5 (timestamp-only $\alpha$ )	94.7%	97.3%	Optimal
$\alpha+\beta$ (+ relevance)	86.8%	96.0%	Slight Drop
$\alpha+\beta+\gamma$ (+ usage)	10.5%	18.9%	Severe Crash
Full Multi-Factor (+ stability)	10.5%	18.9%	Severe Crash

Table 4: Ablation demonstrating the mathematical failure of non-temporal multi-factor scoring.

Figure 4: Relationship between mean number of chunks flagged as shadowed by

\kappa

and the accuracy change induced by GC-Mem. More removable shadows correlate with larger lifts when the detector is well calibrated.

As demonstrated in Table 4, utilizing the pure temporal operator ( $\Phi_{\mathcal{T}}$ ) yields 94.7% CR-Acc. However, introducing usage frequency ( $\gamma$ ) causes accuracy to crash to 10.5%. Because stale chunks have resided in memory longer, they inherently accumulate higher usage and stability scores. Multi-factor scoring systematically assigns higher dominance to the exact contradictions that must be removed, mathematically violating Theorem 4.1. Temporal dominance is uniquely required for state convergence.

6.5 Bag-of-Facts Evaluation: Isolating Protocol Ceiling

To decouple the fundamental efficacy of the $\Phi_{\mathcal{T}}$ operator from the variance of dense retrieval noise, we conducted a "Bag-of-Facts" ablation. In this setting, the retrieval bottleneck is entirely removed ( $k\to\infty$ ), and all relevant historical and updated chunks are placed directly into the context window.

Model	Mean Std. RAG	Mean GC-Mem	$\Delta$ GC Lift	Shadows Pruned
DeepSeek-V3	11.9%	54.3%	+42.4p	13.28
GPT-4o	12.9%	53.3%	+40.3p	15.70
Phi-4	12.3%	44.6%	+32.3p	13.73
04-mini	19.2%	33.4%	+14.2p	6.21

Table 5: Models ranked by pooled mean

\Delta

GC across dataset sizes in the Bag-of-Facts setting. Removing the retrieval limit exposes the immense baseline vulnerability to the Majority Vote Trap (averaging 12% accuracy). GC-Mem consistently excises 13-15 obsolete chunks per query, recovering the state.

As shown in Table 5, when standard RAG is forced to reason over the entire contradictory timeline, its accuracy collapses to roughly 12%. The Majority Vote Trap overwhelms the attention mechanism entirely. However, GC-Mem dynamically prunes an average of 13.28 shadowed chunks per query on DeepSeek-V3, lifting generation accuracy by a massive +42.4 percentage points. This confirms that GC-Mem scales proportionally with retrieval volume: as contexts become more dense with contradictions, the necessity of explicit temporal pruning becomes critical.

6.6 Computational Overhead and Latency

In production architectures, maintaining strict latency bounds is essential. The pairwise conflict detection $\kappa$ operates via a lightweight NLI prompt strictly bounded by $k$ , running exclusively over the retrieved candidate set rather than the global database.

Dataset	STM Buffer	Mean Wall Time (s)	Per-Instance Latency (s)	Avg. Shadows
Medium	STM=off	525s	10.9s	1.55
Medium	STM=on	1276s	26.6s	3.52
Large	STM=off	685s	10.5s	1.39
Large	STM=on	973s	15.0s	4.07

Table 6: End-to-end wall time and per-instance latency by dataset and STM setting. STM increases cost because the unioned candidate set is larger, requiring

\kappa

to evaluate more conflict pairs, yet it yields higher global accuracy by pruning more obsolete shadows (up to 4.07 per instance).

Table 6 details the end-to-end latency. For a standard $k=10$ retrieval without an STM buffer, the GC-Mem protocol operates efficiently at 10.5 seconds per instance. The introduction of an STM buffer naturally inflates the candidate set, thereby increasing the $O(k^{2})$ $\kappa$ evaluation permutations (elevating latency to 15.0s). However, this overhead remains substantially lower than orchestrating multi-turn, generative LLM self-reflection loops, establishing GC-Mem as a highly viable inference-time architecture.

7 Discussion and Limitations

When NOT to use GC-Mem. Our precision/recall analysis dictates a strict deployment guideline: GC-Mem should only be implemented when paired with a conflict detector capable of $>50\%$ recall. Models like gpt-4.1-mini exhibit 100% precision but $<1\%$ recall; they fail to detect contradictions, rendering the protocol inert. Furthermore, partial detection (as seen with gpt-4o) can inadvertently leave the context mixed, confusing the generator further.

Limitations. While the benchmark limits keyword leakage, it remains a synthetic environment optimized for single-attribute mutations. Future work must validate GC-Mem on live, concurrent multi-entity mutations utilizing organic interaction logs. Additionally, while the $O(k^{2})$ conflict check overhead is modest for typical $k\in[5,50]$ (adding $\sim 250$ ms to a standard $400$ ms RAG pipeline), it may prove prohibitive for ultra-low latency requirements.

8 Conclusion

This work proves that semantic shadowing fatally compromises standard RAG architectures over long horizons. By establishing the theoretical bounds of the Majority Vote Trap, we demonstrate that heuristic time-decay and multi-factor scoring both fail to preserve state consistency. We offer GC-Mem, utilizing a pure temporal dominance operator ( $\Phi_{\mathcal{T}}$ ), which empirically restores $>90\%$ accuracy and provides a mathematically sound protocol for mutable RAG systems.

References

Dhingra et al. [2022] Bhuwan Dhingra, Jeremy Cole, Julian Martin Eisenschlos, Dan Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 2022.
Edge et al. [2024] Darren Edge, Ha Trinh, Newman Cheng, et al. Graphrag: Unlocking llm discovery on narrative private data. arXiv preprint arXiv:2404.16130, 2024.
Gutiérrez et al. [2024] Bernal Jiménez Gutiérrez et al. Hipporag: Neurobiologically inspired long-term memory for large language models. In NeurIPS, 2024.
Karpov et al. [2024] Andrei Karpov, Yiyang Li, Aman Madaan, Danqi Chen, and Hannaneh Hajishirzi. A survey of retrieval-augmented generation for large language models. arXiv preprint arXiv:2410.12837, 2024.
Levy et al. [2024] Itay Levy, Michal Moshkovitz, and Reut Tsarfaty. Safe and effective long-context reasoning in code generation. arXiv preprint arXiv:2407.13912, 2024.
Lewis et al. [2020] Patrick Lewis et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In NeurIPS, 2020.
Liska et al. [2022] Adam Liska, Arun Chaganty, Danqi Chen, and Jason Weston. Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models. In International Conference on Machine Learning, 2022.
Liu et al. [2023] Nelson F Liu et al. Lost in the middle: How language models use long contexts. TACL, 2023.
Meng et al. [2022] Kevin Meng, Arnab Sharma, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. In Advances in Neural Information Processing Systems, 2022.
Meng et al. [2023] Kevin Meng, David Sen, Dan Friedman, Yihan Pan, and Yonatan Belinkov. Mass-editing memory in a transformer. In International Conference on Learning Representations, 2023.
Packer et al. [2023] Charles Packer, Vivian Fang, Shishir Patil, Kevin Lin, and Joseph E. Gonzalez. Memgpt: Towards llms as operating systems. arXiv preprint arXiv:2310.08560, 2023.
Park et al. [2023] Joon Sung Park et al. Generative agents: Interactive simulacra of human behavior. arXiv, 2023.
Shinn et al. [2023] Noah Shinn, Federico Cassano, Eli Berman, Ashwin Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. arXiv preprint arXiv:2303.11366, 2023.
Wang et al. [2023] Guanzhi Wang, Yuqi Xie, Wenhao Jiang, Ajay Mandlekar, Haoqi Fan, Anima Anandkumar, and Yuke Wang. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023.
Wang et al. [2025] Z Wang et al. Evaluating memory in llm agents via incremental multi-turn interactions. OpenReview, 2025.
Zhang et al. [2025] Y Zhang et al. Rag-critic: Leveraging automated critic-guided agentic workflow for retrieval augmented generation. In ACL, 2025.

Appendix A Artifact-Backed Results Appendix

This appendix collects the full experimental tables and figures underlying the main text: dataset statistics, headline aggregates, per-condition breakdowns, ablations, and supplementary plots. Together they document every benchmark cell, including retrieval recalls, conflict removals, latency, and significance summaries.

A.1 Dataset Construction and Composition

Table A.1: GCMEMv2 chat dataset statistics. Per-instance means and standard deviations are reported for the number of old, new and distractor turns; "Total turns" is the dataset-wide token/turn count and "Gen. model" indicates the LLM used to generate the synthetic sessions.

Dataset	$N$	Entity	Severity	$n_{\mathrm{old}}$	$m_{\mathrm{new}}$	$n_{\mathrm{dist}}$	Total turns	Gen. model
		types	tiers	mean (SD)	mean (SD)	mean (SD)
Chat-Medium (n=48)	48	12	4	27.5 (14.9)	3.0 (1.2)	12.5 (1.7)	2,033	GPT-4o
Chat-Large (n=65)	65	13	5	27.0 (13.4)	3.0 (1.1)	11.4 (1.7)	2,690	GPT-4o

Figure A.1: Distribution of synthetic entity attributes (stacked by contradiction severity) in Chat-Medium (

n{=}48

) and Chat-Large (

n{=}65

Figure A.2: Per-instance counts of chat turns by scenario role: old (establishes superseded state), new (current ground truth), and distractor (neutral context).

Figure A.3: Illustrative multi-turn instance: user/assistant turns are color-coded by role to show how state change is conveyed without explicit status keywords.

A.2 Primary Benchmark Tables

Dataset	STM	Std. RAG	Timestamp	GC-Mem	$\Delta$ GC
Medium	STM=off	24.4%	24.4%	40.2%	+15.8p
Medium	STM=on	46.1%	71.6%	68.1%	+22.0p
Large	STM=off	15.9%	16.2%	30.4%	+14.4p
Large	STM=on	34.4%	67.5%	61.5%	+27.1p

Table A.2: Headline accuracy (%) averaged across models and prompt modes for each dataset×STM condition. Columns compare Standard RAG, Timestamp re-ranking, and GC-Mem;

\Delta

GC reports the GC-Mem improvement in percentage points. STM=on rows indicate the Short-Term Memory recency buffer was enabled.

Table A.3: Per-condition accuracy breakdown across models, prompt modes, STM settings and datasets. Columns: Std = Standard RAG, TS = Timestamp re-rank baseline, GC = GC-Mem (best shown in bold).

\Delta

GC = GC-Mem

-

Standard RAG (percentage points). Rec_old/new/dis = retrieval recall for old-state / new-state / distractor chunks. "Shadows" is mean number of chunks removed by the conflict oracle

\kappa

per instance; "Err" lists API errors excluded from accuracy; "Sec" shows mean seconds per instance. This granular table is the source for the per-condition figures in the report.

Dataset	STM	Model	Mode	Std	TS	GC	$\Delta$ GC	Rec_old	Rec_new	Rec_dis	Shadows	$n$	Err	Sec
Medium	STM=off	DeepSeek-V3	Forced-choice	25.0%	20.8%	45.8%	+20.8p	24.1%	37.2%	5.0%	1.73	48	0	267
	STM=off	DeepSeek-V3	Allow-unclear	23.4%	19.1%	40.4%	+17.0p	24.3%	38.0%	5.0%	1.77	47	1	259
	STM=off	GPT-4o	Forced-choice	20.8%	22.9%	47.9%	+27.1p	24.1%	37.2%	5.0%	2.27	48	0	669
	STM=off	GPT-4o	Allow-unclear	12.5%	18.8%	29.2%	+16.7p	24.1%	37.2%	5.0%	2.19	48	0	486
	STM=off	GPT-4.1	Forced-choice	20.8%	22.9%	43.8%	+22.9p	24.1%	37.2%	5.0%	1.73	48	0	387
	STM=off	GPT-4.1	Allow-unclear	22.9%	22.9%	35.4%	+12.5p	24.1%	37.2%	5.0%	1.69	48	0	375
	STM=off	GPT-4.1-mini	Forced-choice	33.3%	31.2%	33.3%	+0.0p	24.1%	37.2%	5.0%	0.92	48	0	561
	STM=off	GPT-4.1-mini	Allow-unclear	29.2%	27.1%	37.5%	+8.3p	24.1%	37.2%	5.0%	0.88	48	0	560
	STM=off	GPT-5-Chat	Forced-choice	34.0%	29.8%	48.9%	+14.9p	23.8%	38.0%	5.1%	1.70	47	1	239
	STM=off	GPT-5-Chat	Allow-unclear	23.4%	25.5%	40.4%	+17.0p	24.2%	38.0%	4.9%	1.66	47	1	243
	STM=off	o4-mini	Forced-choice	25.0%	25.0%	39.6%	+14.6p	24.1%	37.2%	5.0%	0.92	48	0	1347
	STM=off	o4-mini	Allow-unclear	22.9%	18.8%	33.3%	+10.4p	24.1%	37.2%	5.0%	0.90	48	0	1330
	STM=off	Phi-4	Forced-choice	31.2%	35.4%	47.9%	+16.7p	24.1%	37.2%	5.0%	1.67	48	0	330
	STM=off	Phi-4	Allow-unclear	16.7%	20.8%	39.6%	+22.9p	24.1%	37.2%	5.0%	1.73	48	0	297
	STM=on	DeepSeek-V3	Forced-choice	37.5%	79.2%	70.8%	+33.3p	32.1%	98.2%	43.3%	3.81	48	0	868
	STM=on	DeepSeek-V3	Allow-unclear	29.2%	62.5%	60.4%	+31.2p	32.1%	98.2%	43.3%	3.77	48	0	914
	STM=on	GPT-4o	Forced-choice	43.8%	79.2%	77.1%	+33.3p	32.1%	98.2%	43.3%	4.77	48	0	1198
	STM=on	GPT-4o	Allow-unclear	45.8%	62.5%	68.8%	+22.9p	32.1%	98.2%	43.3%	4.71	48	0	1252
	STM=on	GPT-4.1	Forced-choice	46.8%	76.6%	70.2%	+23.4p	32.2%	98.2%	43.0%	4.79	47	1	733
	STM=on	GPT-4.1	Allow-unclear	40.4%	72.3%	68.1%	+27.7p	32.2%	98.2%	43.0%	4.91	47	1	720
	STM=on	GPT-4.1-mini	Forced-choice	58.3%	72.9%	66.7%	+8.3p	32.1%	98.2%	43.3%	1.90	48	0	1514
	STM=on	GPT-4.1-mini	Allow-unclear	52.1%	72.9%	62.5%	+10.4p	32.1%	98.2%	43.3%	2.17	48	0	1092
	STM=on	GPT-5-Chat	Forced-choice	53.2%	72.3%	72.3%	+19.1p	32.2%	98.2%	43.0%	3.94	47	1	1194
	STM=on	GPT-5-Chat	Allow-unclear	53.2%	68.1%	70.2%	+17.0p	32.2%	98.2%	43.0%	3.91	47	1	1060
	STM=on	o4-mini	Forced-choice	45.8%	68.8%	58.3%	+12.5p	32.1%	98.2%	43.3%	1.44	48	0	3113
	STM=on	o4-mini	Allow-unclear	41.7%	62.5%	56.2%	+14.6p	32.1%	98.2%	43.3%	1.50	48	0	2874
	STM=on	Phi-4	Forced-choice	58.3%	85.4%	83.3%	+25.0p	32.1%	98.2%	43.3%	3.92	48	0	709
	STM=on	Phi-4	Allow-unclear	39.6%	66.7%	68.8%	+29.2p	32.1%	98.2%	43.3%	3.71	48	0	623
Large	STM=off	DeepSeek-V3	Forced-choice	15.4%	13.9%	38.5%	+23.1p	22.7%	29.9%	4.5%	1.54	65	0	366
	STM=off	DeepSeek-V3	Allow-unclear	15.5%	19.0%	31.0%	+15.5p	22.9%	29.3%	4.6%	1.38	58	7	365
	STM=off	GPT-4o	Forced-choice	15.4%	16.9%	35.4%	+20.0p	22.7%	29.9%	4.5%	2.34	65	0	459
	STM=off	GPT-4o	Allow-unclear	7.7%	7.7%	30.8%	+23.1p	22.7%	29.9%	4.5%	2.32	65	0	441
	STM=off	GPT-4.1	Forced-choice	15.4%	16.9%	33.9%	+18.5p	22.7%	29.9%	4.5%	1.95	65	0	480
	STM=off	GPT-4.1	Allow-unclear	16.9%	21.5%	29.2%	+12.3p	22.7%	29.9%	4.5%	1.82	65	0	478
	STM=off	GPT-4.1-mini	Forced-choice	18.5%	21.5%	27.7%	+9.2p	22.7%	29.9%	4.5%	0.60	65	0	414
	STM=off	GPT-4.1-mini	Allow-unclear	20.0%	18.5%	21.5%	+1.5p	22.7%	29.9%	4.5%	0.62	65	0	407
	STM=off	GPT-5-Chat	Forced-choice	20.0%	16.0%	36.0%	+16.0p	22.4%	29.9%	5.8%	1.84	50	15	650
	STM=off	GPT-5-Chat	Allow-unclear	20.0%	15.4%	35.4%	+15.4p	22.7%	29.9%	4.5%	1.51	65	0	826
	STM=off	o4-mini	Forced-choice	10.8%	13.9%	27.7%	+16.9p	22.7%	29.9%	4.5%	0.55	65	0	1895
	STM=off	o4-mini	Allow-unclear	12.3%	12.3%	23.1%	+10.8p	22.7%	29.9%	4.5%	0.60	65	0	2139
	STM=off	Phi-4	Forced-choice	18.5%	20.0%	30.8%	+12.3p	22.7%	29.9%	4.5%	1.15	65	0	345
	STM=off	Phi-4	Allow-unclear	16.9%	13.9%	24.6%	+7.7p	22.7%	29.9%	4.5%	1.23	65	0	331
	STM=on	DeepSeek-V3	Forced-choice	27.7%	80.0%	69.2%	+41.5p	29.1%	98.5%	47.8%	4.51	65	0	1270
	STM=on	DeepSeek-V3	Allow-unclear	23.1%	61.5%	63.1%	+40.0p	29.1%	98.5%	47.8%	4.46	65	0	1207
	STM=on	GPT-4o	Forced-choice	36.9%	80.0%	83.1%	+46.2p	29.1%	98.5%	47.8%	5.51	65	0	1593
	STM=on	GPT-4o	Allow-unclear	33.9%	47.7%	58.5%	+24.6p	29.1%	98.5%	47.8%	5.43	65	0	1613
	STM=on	GPT-4.1	Forced-choice	28.1%	79.7%	67.2%	+39.1p	29.3%	98.4%	48.2%	5.09	64	1	794
	STM=on	GPT-4.1	Allow-unclear	33.3%	55.6%	55.6%	+22.2p	28.9%	98.4%	48.0%	4.87	63	2	791
	STM=on	GPT-4.1-mini	Forced-choice	36.9%	75.4%	55.4%	+18.5p	29.1%	98.5%	47.8%	2.29	65	0	1019
	STM=on	GPT-4.1-mini	Allow-unclear	35.4%	60.0%	47.7%	+12.3p	29.1%	98.5%	47.8%	2.20	65	0	946
	STM=on	GPT-5-Chat	Forced-choice	39.7%	76.2%	68.2%	+28.6p	28.9%	98.4%	48.0%	4.25	63	2	1166
	STM=on	GPT-5-Chat	Allow-unclear	39.1%	68.8%	65.6%	+26.6p	29.3%	98.4%	48.2%	4.31	64	1	982
	STM=on	o4-mini	Forced-choice	33.9%	70.8%	44.6%	+10.8p	29.1%	98.5%	47.8%	1.62	65	0	3954
	STM=on	o4-mini	Allow-unclear	32.3%	58.5%	46.2%	+13.9p	29.1%	98.5%	47.8%	1.54	65	0	4030
	STM=on	Phi-4	Forced-choice	47.7%	80.0%	75.4%	+27.7p	29.1%	98.5%	47.8%	4.28	65	0	762
	STM=on	Phi-4	Allow-unclear	33.9%	50.8%	61.5%	+27.7p	29.1%	98.5%	47.8%	4.42	65	0	792

STM	Model	Forced-choice $\Delta$ GC	Allow-unclear $\Delta$ GC
STM=off	DeepSeek-V3	+22.0p	+16.3p
STM=off	GPT-4o	+23.5p	+19.9p
STM=off	GPT-4.1	+20.7p	+12.4p
STM=off	GPT-4.1-mini	+4.6p	+4.9p
STM=off	GPT-5-Chat	+15.4p	+16.2p
STM=off	o4-mini	+15.8p	+10.6p
STM=off	Phi-4	+14.5p	+15.3p
STM=on	DeepSeek-V3	+37.4p	+35.6p
STM=on	GPT-4o	+39.7p	+23.8p
STM=on	GPT-4.1	+31.2p	+24.9p
STM=on	GPT-4.1-mini	+13.4p	+11.4p
STM=on	GPT-5-Chat	+23.9p	+21.8p
STM=on	o4-mini	+11.6p	+14.2p
STM=on	Phi-4	+26.3p	+28.4p

Table A.4: Prompt-mode ablation: mean

\Delta

GC (pp) across both datasets for Forced-choice vs. Allow-unclear prompts, per model and STM setting. Bold indicates the stronger prompt mode. GC-Mem improves over Standard RAG in all 12 cells.

Dataset	STM	Recall_old	Recall_new	Recall_dis	Cells
Medium	STM=off	24.1%	37.4%	5.0%	14/6
Medium	STM=on	32.1%	98.2%	43.2%	14/6
Large	STM=off	22.7%	29.9%	4.6%	14/6
Large	STM=on	29.1%	98.5%	47.9%	14/6

Table A.5: Retrieval recall before and after activating the Short-Term Memory (STM) buffer, averaged over 3 models

\times

2 prompt modes. Recall_new = fraction of updated (new-truth) chunks retrieved; Recall_dis = fraction of distractor chunks retrieved (lower is better). STM raises Recall_new from

\sim

37% to

\sim

98%.

STM	Model	Mode	Avg. Shadows Removed	Datasets
STM=off	DeepSeek-V3	Forced-choice	1.64	2/2
STM=off	DeepSeek-V3	Allow-unclear	1.57	2/2
STM=off	GPT-4o	Forced-choice	2.30	2/2
STM=off	GPT-4o	Allow-unclear	2.25	2/2
STM=off	GPT-4.1	Forced-choice	1.84	2/2
STM=off	GPT-4.1	Allow-unclear	1.75	2/2
STM=off	GPT-4.1-mini	Forced-choice	0.76	2/2
STM=off	GPT-4.1-mini	Allow-unclear	0.75	2/2
STM=off	GPT-5-Chat	Forced-choice	1.77	2/2
STM=off	GPT-5-Chat	Allow-unclear	1.58	2/2
STM=off	o4-mini	Forced-choice	0.74	2/2
STM=off	o4-mini	Allow-unclear	0.75	2/2
STM=off	Phi-4	Forced-choice	1.41	2/2
STM=off	Phi-4	Allow-unclear	1.48	2/2
STM=on	DeepSeek-V3	Forced-choice	4.16	2/2
STM=on	DeepSeek-V3	Allow-unclear	4.12	2/2
STM=on	GPT-4o	Forced-choice	5.14	2/2
STM=on	GPT-4o	Allow-unclear	5.07	2/2
STM=on	GPT-4.1	Forced-choice	4.94	2/2
STM=on	GPT-4.1	Allow-unclear	4.89	2/2
STM=on	GPT-4.1-mini	Forced-choice	2.09	2/2
STM=on	GPT-4.1-mini	Allow-unclear	2.19	2/2
STM=on	GPT-5-Chat	Forced-choice	4.09	2/2
STM=on	GPT-5-Chat	Allow-unclear	4.11	2/2
STM=on	o4-mini	Forced-choice	1.53	2/2
STM=on	o4-mini	Allow-unclear	1.52	2/2
STM=on	Phi-4	Forced-choice	4.10	2/2
STM=on	Phi-4	Allow-unclear	4.06	2/2

Table A.6: Mean number of shadow chunks removed per instance by the GC-Mem operator

\kappa

, averaged over both datasets. STM=on substantially increases the number of outdated chunks surfaced and pruned, confirming that STM provides richer context for the conflict detector.

Dataset	Model	Mode	Std. RAG	Timestamp	GC-Mem	Winner	Gap
Medium	DeepSeek-V3	Forced-choice	37.5%	79.2%	70.8%	Timestamp	+8.3pp
	DeepSeek-V3	Allow-unclear	29.2%	62.5%	60.4%	Timestamp	+2.1pp
	GPT-4o	Forced-choice	43.8%	79.2%	77.1%	Timestamp	+2.1pp
	GPT-4o	Allow-unclear	45.8%	62.5%	68.8%	GC-Mem	+6.2pp
	GPT-4.1	Forced-choice	46.8%	76.6%	70.2%	Timestamp	+6.4pp
	GPT-4.1	Allow-unclear	40.4%	72.3%	68.1%	Timestamp	+4.3pp
	GPT-4.1-mini	Forced-choice	58.3%	72.9%	66.7%	Timestamp	+6.2pp
	GPT-4.1-mini	Allow-unclear	52.1%	72.9%	62.5%	Timestamp	+10.4pp
	GPT-5-Chat	Forced-choice	53.2%	72.3%	72.3%	Tie	0.0pp
	GPT-5-Chat	Allow-unclear	53.2%	68.1%	70.2%	GC-Mem	+2.1pp
	o4-mini	Forced-choice	45.8%	68.8%	58.3%	Timestamp	+10.4pp
	o4-mini	Allow-unclear	41.7%	62.5%	56.2%	Timestamp	+6.2pp
	Phi-4	Forced-choice	58.3%	85.4%	83.3%	Timestamp	+2.1pp
	Phi-4	Allow-unclear	39.6%	66.7%	68.8%	GC-Mem	+2.1pp
Large	DeepSeek-V3	Forced-choice	27.7%	80.0%	69.2%	Timestamp	+10.8pp
	DeepSeek-V3	Allow-unclear	23.1%	61.5%	63.1%	GC-Mem	+1.5pp
	GPT-4o	Forced-choice	36.9%	80.0%	83.1%	GC-Mem	+3.1pp
	GPT-4o	Allow-unclear	33.9%	47.7%	58.5%	GC-Mem	+10.8pp
	GPT-4.1	Forced-choice	28.1%	79.7%	67.2%	Timestamp	+12.5pp
	GPT-4.1	Allow-unclear	33.3%	55.6%	55.6%	Tie	0.0pp
	GPT-4.1-mini	Forced-choice	36.9%	75.4%	55.4%	Timestamp	+20.0pp
	GPT-4.1-mini	Allow-unclear	35.4%	60.0%	47.7%	Timestamp	+12.3pp
	GPT-5-Chat	Forced-choice	39.7%	76.2%	68.2%	Timestamp	+7.9pp
	GPT-5-Chat	Allow-unclear	39.1%	68.8%	65.6%	Timestamp	+3.1pp
	o4-mini	Forced-choice	33.9%	70.8%	44.6%	Timestamp	+26.2pp
	o4-mini	Allow-unclear	32.3%	58.5%	46.2%	Timestamp	+12.3pp
	Phi-4	Forced-choice	47.7%	80.0%	75.4%	Timestamp	+4.6pp
	Phi-4	Allow-unclear	33.9%	50.8%	61.5%	GC-Mem	+10.8pp

Table A.7: Head-to-head comparison of Timestamp re-ranking and GC-Mem under STM=on (where the STM buffer dramatically improves retrieval recall). Bold marks the stronger of the two non-trivial baselines. Timestamp becomes competitive under STM because high recall reduces its recency-ranking advantage gap, yet GC-Mem retains an edge in Allow-unclear (lenient) conditions.

Dataset	STM	Mean Wall Time (s)	Per-instance (s)	Avg. Shadows	Cells
Medium	STM=off	525	10.9	1.55	14/6
Medium	STM=on	1276	26.6	3.52	14/6
Large	STM=off	685	10.5	1.39	14/6
Large	STM=on	1494	23.0	3.91	14/6

Table A.8: Wall-clock time per condition averaged over 3 models

\times

2 prompt modes, and mean shadows removed per instance. STM=on increases latency due to additional STM retrieval and conflict detection, but also removes substantially more shadow chunks.

A.3 Primary Benchmark Figures

Figure A.4: Effect of enabling STM (Redis FIFO of recent chunks) on standard RAG vs. GC-Mem accuracy. STM raises both baselines by improving exposure of new-state evidence; GC-Mem retains a larger gap because it removes contradictions that simple recency heuristics leave in context.

A.4 Example Prompt

Figure A.5: Example prompt templates used for the prompt-mode ablation. The figure shows the two prompt styles compared: Forced-choice (must pick a label) and Allow-unclear (model may respond with “unclear”). These were used across all models to measure robustness of GC-Mem to prompt framing.

Figure A.6: Retrieval recall by chunk type (old-state, new-state, distractor) comparing STM off vs on. STM substantially raises recall of new-state chunks, which increases the number of contradictory pairs in the candidate set and therefore the benefit of applying the temporal dominance operator

\Phi_{\mathcal{T}}

Figure A.8: Forest plot of paired bootstrap 95% confidence intervals for

\Delta

GC (10,000 resamples per cell). Intervals entirely above zero support a consistent gain; intervals crossing zero flag weak

\kappa

on that model or incomplete cells.

Figure A.9:

\Delta

GC broken down by synthetic entity type. Positive bars across categories indicate that GC-Mem is not driven by a single attribute family.

A.5 Bag-of-Facts (No-Retrieval) Ablation

Dataset	$n$	Mean Std	$\sigma$	Mean TS	Mean GC	Mean $\Delta$ GC	$\sigma$ ( $\Delta$ GC)
Medium	48	20.4%	$\pm$ 10.7	18.5%	55.2%	+34.8p	$\pm$ 12.2
Large	65	12.1%	$\pm$ 4.2	11.7%	36.5%	+24.4p	$\pm$ 11.6

Table A.9: Bag-of-facts (k=

\infty

) headline results: mean accuracy and

\Delta

GC across 7 models. Std = Standard RAG (all facts in context without GC), TS = Timestamp re-ranking, GC = GC-Mem.

\sigma

is sample standard deviation across 7 models.

Model	Dataset	Std. RAG	Timestamp	GC-Mem	$\Delta$ GC	Shadows	$n$
DeepSeek-V3	Medium	14.6%	12.5%	56.2%	+41.7p	13.58	48
	Large	9.2%	12.3%	52.3%	+43.1p	12.98	65
GPT-4o	Medium	16.7%	18.8%	60.4%	+43.8p	16.67	48
	Large	9.2%	6.2%	46.2%	+36.9p	14.72	65
GPT-4.1	Medium	21.7%	15.2%	60.9%	+39.1p	13.02	48
	Large	14.1%	9.4%	31.2%	+17.2p	10.47	65
GPT-4.1-mini	Medium	14.6%	18.8%	37.5%	+22.9p	7.42	48
	Large	7.7%	7.7%	24.6%	+16.9p	5.14	65
GPT-5-Chat	Medium	39.1%	21.7%	71.7%	+32.6p	13.39	48
	Large	17.5%	11.1%	44.4%	+27.0p	10.52	65
o4-mini	Medium	29.2%	31.2%	43.8%	+14.6p	7.25	48
	Large	9.2%	15.4%	23.1%	+13.9p	5.17	65
Phi-4	Medium	7.0%	11.6%	55.8%	+48.8p	14.02	48
	Large	17.6%	19.6%	33.3%	+15.7p	13.45	65

Table A.10: Per-model bag-of-facts results across both dataset sizes. Std = Standard RAG accuracy (all facts provided); GC-Mem = accuracy after GC operator prunes outdated shadows (bold).

\Delta

GC = GC-Mem

-

Std. RAG in percentage points. Shadows = mean outdated chunks removed per instance by

\kappa

Rank	Model	Mean Std. RAG	Mean GC-Mem	Mean $\Delta$ GC	Mean Shadows
1	DeepSeek-V3	11.9%	54.3%	+42.4p	13.28
2	GPT-4o	12.9%	53.3%	+40.3p	15.70
3	Phi-4	12.3%	44.6%	+32.3p	13.73
4	GPT-5-Chat	28.3%	58.1%	+29.8p	11.96
5	GPT-4.1	17.9%	46.1%	+28.2p	11.75
6	GPT-4.1-mini	11.1%	31.1%	+19.9p	6.28
7	o4-mini	19.2%	33.4%	+14.2p	6.21

Table A.11: Models ranked by pooled mean

\Delta

GC (average of medium and large datasets). Mean Std. RAG and Mean GC-Mem are averages of the two dataset sizes. Models with lower baseline accuracy (Std) tend to show larger GC-Mem gains, indicating that GC pruning is most effective when the unfiltered context is most misleading.

Model	Chat-RAG no-STM	Chat-RAG STM-on	Bag-of-facts (k= $\infty$ )
DeepSeek-V3	+18.9p	+32.3p	+41.7p
GPT-5-Chat	+16.0p	+18.1p	+32.6p
Phi-4	+19.8p	+27.1p	+48.8p

Table A.12:

\Delta

GC (pp) on the medium dataset under three setups for the three models common to all experiments. Chat-RAG values are averaged over Forced-choice and Allow-unclear prompt modes. Bag-of-facts removes the Qdrant retrieval bottleneck (k=

\infty

, all facts in context) before GC-Mem is applied. The monotone increase Chat-RAG no-STM

<

Chat-RAG STM-on

<

Bag-of-facts confirms that GC-Mem advantage scales with retrieval quality: when the context contains more outdated facts, GC pruning delivers a larger lift.

Figure A.10: Bag-of-facts (no retrieval) setting:

\Delta

GC by model when all old and new memories are placed directly in context (

k{\rightarrow}\infty

). This isolates the value of

\Phi_{\mathcal{T}}

from retrieval noise.

Figure A.11: Comparison of mean

\Delta

GC across chat-RAG (with and without STM) and bag-of-facts. Larger candidate sets with more conflicting statements yield larger GC-Mem lifts.

Figure A.12: Distribution of shadow counts identified by

\kappa

in the bag-of-facts regime, by model.

Figure A.13: Standard RAG accuracy vs.

\Delta

GC in bag-of-facts: weaker baselines often coincide with more headroom for conflict pruning.

Appendix B Mathematical Proofs

B.1 Proof of Theorem 1: Asymptotic Recall Decay

Proof.

Let the total memory stream contain $N$ chunks asserting the stale state $S_{old}$ and $M$ chunks asserting the valid state $S_{new}$ . Under the Accumulation Hypothesis (Assumption 3.3), $N\gg M$ .

We model the retrieval of top- $k$ chunks $\mathcal{R}_{k}$ as sampling without replacement from the population $N+M$ . Let $X$ be the random variable denoting the number of valid chunks ( $S_{new}$ ) successfully retrieved. If retrieval were purely uniform, $X$ would follow a hypergeometric distribution:

P(X=x)=\frac{\binom{M}{x}\binom{N}{k-x}}{\binom{N+M}{k}}

(5)

The expected value is $\mathbb{E}[X]=k\frac{M}{N+M}$ .

However, dense retrievers utilize cosine similarity over high-dimensional embeddings. Under Semantic Equivalence (Lemma 3.2), $|sim(q,S_{old})-sim(q,S_{new})|<\epsilon$ . Because $S_{old}$ constitutes the vast majority of the embedded semantic space, dense retrieval exhibits a structural bias towards the denser centroid of $S_{old}$ phrasings. Therefore, the hypergeometric expected value serves as a strict theoretical upper bound.

Taking the limit as the interaction history scales:

\lim_{N\to\infty}\mathbb{E}[X]\approx\lim_{N\to\infty}k\frac{M}{N+M}=0

(6)

As $N\to\infty$ , the probability of retrieving zero valid evidence ( $P(X=0)$ ) approaches 1, proving Asymptotic Recall Decay. ∎

B.2 Proof of Theorem 2: The Majority Vote Trap

Proof.

Assume the retriever successfully circumvents Theorem 1 and retrieves a mixed context $\mathcal{R}_{k}$ containing $1$ valid chunk ( $S_{new}$ ) and $k-1$ stale chunks ( $S_{old}$ ).

During generation, the LLM utilizes scaled dot-product attention. Let $\alpha_{i}$ denote the attention weight assigned to chunk $c_{i}\in\mathcal{R}_{k}$ . Under Lemma 3.2, the semantic embeddings of $S_{old}$ and $S_{new}$ are equivalent up to $\epsilon<0.10$ . Consequently, the query vector $q$ yields roughly identical dot-products across all $k$ chunks.

Applying the softmax function over these practically uniform logits results in a non-discriminating attention distribution:

\alpha_{i}\approx\frac{1}{k}\quad\forall c_{i}\in\mathcal{R}_{k}

(7)

The final generation probability for a state $y$ is the sum of the attention mass corresponding to the chunks asserting that state.

P(y=S_{old})=\sum_{c_{i}\in S_{old}}\alpha_{i}\approx\frac{k-1}{k}

(8)

P(y=S_{new})=\sum_{c_{j}\in S_{new}}\alpha_{j}\approx\frac{1}{k}

(9)

For any typical retrieval window ( $k\geq 5$ ), $\frac{k-1}{k}\gg\frac{1}{k}$ . The probability mass overwhelmingly favors the stale state. Thus, even when valid evidence is perfectly retrieved, generation collapses due to attention dilution, proving the Majority Vote Trap. ∎

The Immutable Past: Formalizing State Mutability and Conflict Resolution in Mutable RAG

Abstract

1 Introduction

2 Related Work

3 Theoretical Framework

3.1 Problem Formulation

Definition 3.1 (Mutable Entity State & Multi-Transition Dynamics).

Lemma 3.2 (Semantic Equivalence).

Assumption 3.3 (Accumulation Hypothesis (N≫MN\gg M)).

3.2 The Shadowing Failure

Theorem 3.4 (Asymptotic Recall Decay).

Theorem 3.5 (The Majority Vote Trap).

4 The GC-Mem Protocol

4.1 Memory Architecture and Candidate Retrieval

4.2 Conflict Oracle and Temporal Dominance

Theorem 4.1 (State Recovery).

5 Experimental Methodology

5.1 Dataset Scale and Rationale

5.2 The Behavioral Pipeline

6 Results

6.1 Timestamp Baselines vs. GC-Mem

6.2 Short-Term Memory (STM) Dynamics

6.3 Detector Dependency and Deployment Threshold

6.4 Ablation: The Failure of Multi-Factor Dominance

6.5 Bag-of-Facts Evaluation: Isolating Protocol Ceiling

6.6 Computational Overhead and Latency

7 Discussion and Limitations

8 Conclusion

References

Appendix A Artifact-Backed Results Appendix

A.1 Dataset Construction and Composition

A.2 Primary Benchmark Tables

A.3 Primary Benchmark Figures

A.4 Example Prompt

A.5 Bag-of-Facts (No-Retrieval) Ablation

Appendix B Mathematical Proofs

B.1 Proof of Theorem 1: Asymptotic Recall Decay

Proof.

B.2 Proof of Theorem 2: The Majority Vote Trap

Proof.

Assumption 3.3 (Accumulation Hypothesis ( $N\gg M$ )).