LongRecall: Multifactor Memory & Retrieval

• LongRecall is a multifaceted construct defining human free recall by dissociating total recall, recency/primacy balance, and initial recall position with minimal correlation among them.
• It formalizes retrieval dynamics via the Fundamental Law of Memory Recall and leverages a three-stage evaluation framework using fact extraction, candidate narrowing, and entailment verification.
• In computational models and agent systems, LongRecall drives enhanced performance by integrating multi-path retrieval, selective memory consolidation, and mechanisms to mitigate forgetting.

Searching arXiv for papers and context on “LongRecall”. LongRecall denotes a line of inquiry concerned with the completeness, structure, and retrieval dynamics of long-term recall across human memory experiments, long-context LLMs, memory-augmented agents, and recall evaluation frameworks. In the psychological sense, the term is used to describe long-term recall as a non-unitary construct comprising total capacity, serial-position effects, and recall strategy or executive-choice components (Tarnow, 2016). In the evaluation sense, “LongRecall” also names a three-stage framework for robust recall assessment in long-form generated text, designed to decompose answers into self-contained facts, narrow candidate matches through lexical and semantic filtering, and verify alignment through structured entailment checks (Ardestani et al., 20 Aug 2025). Across these usages, a recurring theme is that recall cannot be reduced to a single scalar notion of “memory strength”: both human and machine studies emphasize dissociable mechanisms, trade-offs between fidelity and efficiency, and the importance of structured retrieval or verification (Tarnow, 2016).

1. Human long-term recall as a multifactorial construct

In human free recall, LongRecall is characterized as a composite of at least three separable memory properties: total recall, recency and primacy balance (RPB), and initial recall position (Tarnow, 2016). Using single-factor ANOVA on classic free-recall data, large individual differences were reported across all three properties, with effect sizes $\eta$ ranging from $0.09$ to $0.26$ (Tarnow, 2016). For 10-item lists, the reported values were $\eta = 0.26$ for total recall, $\eta = 0.18$ for RPB, and $\eta = 0.20$ for initial recall; for 40-item lists, they were $\eta = 0.17$, $\eta = 0.14$, and $\eta = 0.094$, respectively (Tarnow, 2016). The paper defines the ANOVA effect size as

$\eta^2 = \frac{SS_\text{between}}{SS_\text{total}}$

where $0.09$0 is the sum of squares between subjects and $0.09$1 is the total sum of squares (Tarnow, 2016).

A central result is that these properties are largely uncorrelated. Total recall versus RPB had low $0.09$2 values, reported as $0.09$3 for 10-2 and $0.09$4 for 40-1, and no strong correlations were observed between initial recall position and either total recall or RPB (Tarnow, 2016). This directly motivated the interpretation that free recall reflects multiple relatively independent components rather than a single latent ability (Tarnow, 2016). The paper further proposes that variance in initial recall position across lists may index executive control, with lower variance correlating with higher total recall; the reported correlations were modest, with $0.09$5 for 10-2 and $0.09$6 for 40-1 (Tarnow, 2016). The maximal benefit associated with this inferred executive control was reported as up to $0.09$7 more items for 10-item lists and $0.09$8 more for 40-item lists (Tarnow, 2016).

The same study reports that individual differences decrease in later recalls after the initial recall, and that there is no evidence for strong clustering into distinct memory “types”; instead, the observed distributions were wide and continuous rather than multimodal (Tarnow, 2016). It also notes clinical relevance for the recency/primacy balance, especially because loss of primacy has diagnostic value in conditions such as Alzheimer’s disease (Tarnow, 2016). Taken together, these findings support the use of LongRecall as a multifaceted construct rather than a unitary faculty (Tarnow, 2016).

2. Formal laws and retrieval dynamics in free recall

A second strand of LongRecall research seeks lawful descriptions of recall dynamics. A prominent result is the “Fundamental Law of Memory Recall,” which models free recall as an associative search process reducible to a deterministic walk on random graphs defined by memory representations (Naim et al., 2019). In that model, if $0.09$9 denotes the number of items successfully encoded and present in memory, the average number of recalled items is predicted to be

$0.26$0

This prediction is parameter-free and was verified in a large-scale crowd-sourced free recall and recognition experiment (Naim et al., 2019).

The model assumes sparse, overlapping neuronal assemblies for item representation and a deterministic transition rule in which the next recalled item is the one with the largest similarity to the current item, except that the process cannot immediately return to the item recalled in the prior step (Naim et al., 2019). The resulting trajectory on the similarity matrix continues until it cycles among previously recalled items, at which point recall terminates (Naim et al., 2019). For biological encoding, the chance of returning to a previously visited item in a single step is given as $0.26$1, and further combinatorial constraints reduce this by a factor of $0.26$2, yielding the square-root law above (Naim et al., 2019).

The experimental design paired free recall with recognition in order to estimate $0.26$3 independently of recall output. From recognition accuracy $0.26$4, the number of items in memory was inferred as

$0.26$5

where $0.26$6 is list length (Naim et al., 2019). Plotting recalled items $0.26$7 against estimated $0.26$8 produced a collapse onto the theoretical curve across list lengths, presentation rates, materials, and individuals (Naim et al., 2019). The paper interprets this as evidence that, once encoding variability is factored out, recall operates according to a stereotyped search process common to all people (Naim et al., 2019).

This law complements rather than negates the multifactorial account above. A plausible implication is that the square-root law characterizes one aggregate regularity of retrieval dynamics, while the large individual differences in total recall, RPB, and initial recall position identify orthogonal or partially orthogonal sources of variation around that regularity (Tarnow, 2016). The two perspectives are therefore compatible with a view of LongRecall in which universal search constraints coexist with stable person-level differences.

3. LongRecall as a structured evaluation framework for long-form text

In generative NLP, LongRecall is also the name of a structured approach for robust recall evaluation in long-form text (Ardestani et al., 20 Aug 2025). The framework is motivated by the claim that completeness is crucial in domains such as medicine and law and in tasks such as list-based question answering, while existing metrics based on lexical overlap or holistic LLM-as-a-Judge prompting are error-prone under paraphrase, alternative names, or long complex outputs (Ardestani et al., 20 Aug 2025).

The framework is explicitly three-stage. First, it decomposes both reference and generated answers into granular, self-contained facts, with coreference resolution applied to make facts explicit (Ardestani et al., 20 Aug 2025). The decomposition can rely on heuristics for lists, segmentation of annotated clauses, or semantic decomposition methods such as FActScore-style atomization (Ardestani et al., 20 Aug 2025). Second, it narrows candidate matches for each reference fact through lexical filtering followed by semantic filtering (Ardestani et al., 20 Aug 2025). The lexical step uses fuzzy Jaccard similarity and Longest Common Subsequence, while the semantic step uses dense embeddings and cosine similarity (Ardestani et al., 20 Aug 2025). Third, it applies structured entailment checking with an LLM, either in one-to-one form or through multiple-choice entailment when several generated facts may jointly cover the reference fact (Ardestani et al., 20 Aug 2025).

Formally, if $0.26$9 is the set of reference facts and $\eta = 0.26$0 the set of generated facts, the candidate set for a reference fact $\eta = 0.26$1 is denoted $\eta = 0.26$2, and recall is defined as

$\eta = 0.26$3

Dataset-level recall is then obtained by averaging instance-level recall scores (Ardestani et al., 20 Aug 2025). The framework is presented as modular and interpretable because each extraction, filtering, and entailment decision can be inspected individually (Ardestani et al., 20 Aug 2025).

The empirical evaluation spans QAMPARI, ExpertQA, and RoMQA (Ardestani et al., 20 Aug 2025). Reported results indicate that LongRecall outperforms lexical baselines such as ARecall and ERecall, as well as holistic LLM-as-a-Judge prompting, especially on split-span and domain-rich settings (Ardestani et al., 20 Aug 2025). On QAMPARI and RoMQA, the reported gains over lexical baselines range from $\eta = 0.26$4 to $\eta = 0.26$5 in F1 across standard and challenging splits, and on ExpertQA the paper reports LongRecall at approximately $\eta = 0.26$6–$\eta = 0.26$7 F1 while ERecall is approximately $\eta = 0.26$8 (Ardestani et al., 20 Aug 2025). The entailment checker based on LLaMA-3.3-70B-Instruct is reported to achieve Fleiss’ kappa of $\eta = 0.26$9–$\eta = 0.18$0 with human annotations (Ardestani et al., 20 Aug 2025). These results frame LongRecall not as a generator or retriever, but as a foundational evaluation primitive for systematic completeness assessment (Ardestani et al., 20 Aug 2025).

4. Recall in LLMs: recognition, retention, and context use

Research on LLMs extends LongRecall from human memory and evaluation into model behavior. Few-shot memory experiments show that recognition, recall, and retention in LLMs are dissociable (Orhan, 2023). Across 19 causal LLMs from BLOOM, OPT, GPT-2, and GPT-Neo families, one exposure was generally sufficient for near-perfect recognition, while near-perfect recall of natural sentences typically required three exposures (Orhan, 2023). Recall for the original examples then dropped steeply over the first 10 training updates with new examples, followed by a more gradual decline, although some examples remained recallable near perfectly even after 100K updates (Orhan, 2023). Recognition was reported as much more robust to interference than recall, and natural language sentences were retained better than structureless stimuli (Orhan, 2023).

This dissociation between recognition and recall parallels the distinction between fast familiarity and deliberate recollection later exploited by retrieval systems for personalized agents. RF-Mem formalizes this with a dual-path retriever in which a top-$\eta = 0.18$1 probe is scored by cosine similarity, transformed into a softmax distribution, and summarized by mean similarity $\eta = 0.18$2 and entropy $\eta = 0.18$3 (Zhang et al., 10 Mar 2026). High familiarity leads to a direct top-$\eta = 0.18$4 path, while low familiarity or high uncertainty activates a recollection path that clusters candidates and updates the query by an $\eta = 0.18$5-mix:

$\eta = 0.18$6

Across benchmarks and corpus scales, RF-Mem is reported to outperform both one-shot retrieval and full-context reasoning under fixed budget and latency constraints (Zhang et al., 10 Mar 2026).

Long-context reasoning models reveal another failure mode: retrieval quality can degrade after a short reasoning span, a phenomenon termed “lost-in-thought” (Whitecross et al., 10 Apr 2026). RecaLLM addresses this by interleaving reasoning with explicit in-context retrieval via recall spans delimited by <|start_recall|> and <|end_recall|>, with constrained decoding that forces copied evidence to be a verbatim substring of the searchable context (Whitecross et al., 10 Apr 2026). On RULER and HELMET, RecaLLM is reported to significantly outperform baselines, with gains persisting up to 128K-token contexts despite training samples of at most 10K tokens (Whitecross et al., 10 Apr 2026).

State-space models exhibit a related limitation. “Recall with Reasoning” (RwR) improves Mamba’s long-context behavior by prepending chain-of-thought summarizations distilled from a teacher model during fine-tuning, without architectural changes (Ma et al., 6 May 2025). On LONGMEMEVAL and HELMET, RwR improved LONGMEMEVAL from $\eta = 0.18$7 to $\eta = 0.18$8 at 10K and from $\eta = 0.18$9 to $\eta = 0.20$0 at 100K, while preserving or slightly improving short-context performance (Ma et al., 6 May 2025). These results suggest that long-context recall in sequence models depends not only on capacity, but on training protocols that explicitly teach the model what to recover and when.

5. Long-term recall in agent memory systems

Agent memory systems operationalize LongRecall as the ability to preserve, retrieve, and update task-relevant information over long horizons. LongMemEval-V2 is a benchmark specifically designed to test whether memory systems help agents become “experienced colleagues” in customized web environments (Wu et al., 12 May 2026). It contains 451 manually curated questions spanning five abilities: static state recall, dynamic state tracking, workflow knowledge, environment gotchas, and premise awareness (Wu et al., 12 May 2026). The benchmark pairs questions with history haystacks containing up to 500 trajectories and 115M tokens and evaluates memory via a context-gathering interface consisting of Insert(h) and Query(q) (Wu et al., 12 May 2026).

Within this benchmark, AgentRunbook-R maintains knowledge pools for raw state slices, state transition events, and procedure or hint notes, reaching $\eta = 0.20$1 on the small setting and $\eta = 0.20$2 on the medium setting with latency of approximately 26 seconds per query (Wu et al., 12 May 2026). AgentRunbook-C stores trajectories as files and uses a coding agent to gather evidence in a sandbox, reaching $\eta = 0.20$3 on the small setting and $\eta = 0.20$4 on the medium setting with latency of approximately 108–140 seconds (Wu et al., 12 May 2026). The benchmark-level average accuracy reported for AgentRunbook-C is $\eta = 0.20$5, compared with $\eta = 0.20$6 for the strongest RAG baseline and $\eta = 0.20$7 for the off-the-shelf coding agent baseline (Wu et al., 12 May 2026).

Other systems emphasize different aspects of long-term recall. Engram uses a bi-temporal memory engine with a fast write path for lossless episodes and an asynchronous path that extracts atomic $\eta = 0.20$8 facts into a bi-temporal knowledge graph with supersession chains and provenance (Wang, 5 Jun 2026). On the full 500-question LongMemEval_S, its lean hybrid configuration is reported at $\eta = 0.20$9 accuracy versus $\eta = 0.17$0 for a full-context baseline, with approximately 9.6k retrieved tokens instead of 79k and a McNemar exact test of $\eta = 0.17$1 (Wang, 5 Jun 2026). T-Mem addresses a different gap by introducing descriptive and associative triggers at both item and scene levels, so that memories remain reachable both through surface-similar and latent-semantic cues (Guo et al., 13 Jun 2026). It reports state-of-the-art performance on LoCoMo and LoCoMo-Plus, with overall LoCoMo accuracy of $\eta = 0.17$2 and a LoCoMo-to-LoCoMo-Plus gap of $\eta = 0.17$3, compared with a gap of $\eta = 0.17$4 for HyperMem (Guo et al., 13 Jun 2026). Dual-trace encoding similarly enriches recall by pairing fact traces with scene traces; on LongMemEval-S it is reported at $\eta = 0.17$5 overall accuracy versus $\eta = 0.17$6 for fact-only memory, with particularly large gains on temporal reasoning, knowledge-update tracking, and multi-session aggregation (Stern et al., 14 Apr 2026).

Memory construction and eviction also matter. RecMem introduces a subconscious memory layer indexed by lightweight embeddings and triggers LLM-based consolidation only when sustained recurrence is detected, reporting up to $\eta = 0.17$7 reduction in memory construction token cost while exceeding the accuracy of three state-of-the-art systems (Dai et al., 15 May 2026). LRE approaches the problem from the eviction side with a few-kilobyte logistic-regression scorer over causal features that preserves verbatim “load-bearing” units, matching the overall accuracy of keeping the entire history on agents and exceeding it by $\eta = 0.17$8 on the simplest tasks while reducing peak context size by up to $\eta = 0.17$9 (Lia et al., 18 Jun 2026). These systems indicate that LongRecall in agents is not equivalent to storing everything; it depends on evidence selection, update handling, and retrieval fidelity.

6. Forgetting, continual learning, and long-horizon retention

A major theme in LongRecall research is that durable recall requires explicit mechanisms against forgetting. In personalized agents, Memora benchmarks remembering, reasoning, and recommending over weeks to months of simulated user conversations and introduces Forgetting-Aware Memory Accuracy (FAMA) (Uddin et al., 21 Apr 2026). FAMA is defined as

$\eta = 0.14$0

with

$\eta = 0.14$1

where MPA is Memory Presence Accuracy and FAA is Forgetting Absence Accuracy (Uddin et al., 21 Apr 2026). The metric is designed to penalize reliance on obsolete or invalidated memory, and evaluations of four LLMs and six memory agents show frequent reuse of invalid memories and only marginal gains from existing memory agents (Uddin et al., 21 Apr 2026). The reported score reductions under FAMA range from 10 to 43 points across tasks and durations (Uddin et al., 21 Apr 2026).

Continual-learning studies reach similar conclusions. “Scalable Recollections for Continual Lifelong Learning” replaces raw experience storage with discrete VAE codes stored in an index buffer, reducing per-experience storage from $\eta = 0.14$2 bits to $\eta = 0.14$3 bits with

$\eta = 0.14$4

and total storage $\eta = 0.14$5 bits (Riemer et al., 2017). Under matched storage budgets, it reports substantial retention gains over GEM and replay with raw data, including $\eta = 0.14$6 versus $\eta = 0.14$7 on MNIST-Rotations and $\eta = 0.14$8 versus $\eta = 0.14$9 on CIFAR-100 for GEM-based comparisons (Riemer et al., 2017). In semantic parsing, TotalRecall combines diversified logical-form selection with fast-slow continual learning, reaching within roughly 3% of Oracle retraining while providing a 3–6 times speedup compared to retraining from scratch (Li et al., 2021).

Embodied-action systems show the same stability–plasticity dilemma. In vision-language-action models, uncertainty-guided active data collection improves adaptation efficiency, but fine-tuning only on active recovery data causes catastrophic forgetting (Karli et al., 22 Jun 2026). The best reported configuration, active online collection with full replay, achieved $\eta = 0.094$0 overall, $\eta = 0.094$1 on collected tasks, and $\eta = 0.094$2 on retained tasks, whereas active online new-only training collapsed to $\eta = 0.094$3 overall and $\eta = 0.094$4 on retained tasks (Karli et al., 22 Jun 2026). EWC and low learning rates reduced forgetting only partially and impaired adaptation (Karli et al., 22 Jun 2026). This suggests that LongRecall under continual update is inseparable from memory replay, consolidation, or explicit supersession management.

A related practical observation appears in high-recall extraction from long documents. L3X addresses “long object list extraction from long documents” through a recall-oriented generation stage followed by precision-oriented scrutinization (Singhania et al., 2024). Stage 1 recall rises from approximately $\eta = 0.094$5 for LLM-only baselines to approximately $\eta = 0.094$6–$\eta = 0.094$7 for retrieval-augmented configurations, with R@P50 around $\eta = 0.094$8–$\eta = 0.094$9 after scrutiny (Singhania et al., 2024). Although this setting is extraction rather than memory per se, it illustrates the same general principle: maximizing LongRecall often requires an explicit first-stage design for exhaustive coverage, followed by downstream verification or pruning (Singhania et al., 2024).

7. Conceptual unities and open directions

Across human memory, model evaluation, long-context reasoning, and agent architectures, LongRecall consistently denotes more than raw retrieval volume. In human free recall, it refers to a construct comprising total recall, serial-position balance, and executive-choice effects that are large in individual differences yet relatively uncorrelated (Tarnow, 2016). In recall evaluation, it denotes a pipeline that breaks completeness into fact extraction, candidate narrowing, and entailment verification (Ardestani et al., 20 Aug 2025). In LLMs and agents, it refers to long-horizon retention and retrieval under interference, budget constraints, contradictions, and evolving state (Orhan, 2023).

Several recurring design patterns emerge. First, structured decomposition is central: free recall is broken into dissociable properties, answers into self-contained facts, and agent histories into episodes, facts, traces, or triggers (Tarnow, 2016). Second, retrieval benefits from multi-path or staged processing rather than a single similarity lookup, as seen in LongRecall evaluation, RF-Mem, RecaLLM, T-Mem, L3X, and AgentRunbook (Ardestani et al., 20 Aug 2025). Third, forgetting is not merely failure to retrieve but often failure to manage mutation, supersession, or rehearsal, motivating FAMA, bi-temporal invalidation, replay, recurrence-based consolidation, and learned eviction (Uddin et al., 21 Apr 2026). Fourth, lean, selective memory can outperform full-history replay, as shown by Engram’s retrieved slice beating a full-context baseline on LongMemEval_S (Wang, 5 Jun 2026).

A plausible synthesis is that LongRecall is best understood as a systems property of faithful, context-sensitive, and mutation-aware recovery. On that view, the central challenge is not only whether relevant information exists somewhere in memory, history, or generated text, but whether it remains reachable, attributable, non-obsolete, and sufficiently verified for the downstream task. The papers surveyed here collectively define that challenge across psychology, evaluation, and machine intelligence, while also showing that no single mechanism—recency, lexical overlap, dense similarity, or full-context replay—is sufficient on its own (Tarnow, 2016).