智能体应该记住什么？在有限记忆评估中将记忆保留与检索解耦

# What Should an Agent Remember? Disentangling Retention fromRetrieval in Bounded-Memory Evaluation
Source: [https://arxiv.org/html/2610.00366](https://arxiv.org/html/2610.00366)
## What Should an Agent Remember? Disentangling Retention from Retrieval in Bounded\-Memory EvaluationThanks:Code:[https://github\.com/TheClassicTechno/MemoryLLMAgentProject](https://github.com/TheClassicTechno/MemoryLLMAgentProject)

###### Abstract

A persistent agent must decide both*what to retain*as information arrives and*what to surface*once a query appears, different decisions made at different times with different information, yet memory evaluations can end up comparing methods that differ on both at once\. We build a streaming\-recall environment crossing a retention rule with a selection rule, scoring the same 300 seeded episodes in every cell\. Holding access fixed, query\-aware selection buys\+15\.5\+15\.5percentage points \(pp\) of required\-fact recall \(95% CI\[\+12\.8,\+18\.2\]\[\+12\.8,\+18\.2\]\), whereas the mixed\-cell contrast used by an earlier version of this evaluation reports\+68\.7\+68\.7pp\[\+65\.2,\+72\.0\]\[\+65\.2,\+72\.0\]; the extra\+53\.2\+53\.2pp is access\. With query\-aware, dense, or oracle selection, recall in every bounded retention condition then exactly equals its retention rate: a 768\-dimensional dense retriever moves it by\+0\.00\+0\.00pp\[\+0\.00,\+0\.00\]\[\+0\.00,\+0\.00\], in the bounded recency\-retention condition all 319 observed retrieval failures are evictions rather than ranking errors, and recall there falls to0%0\\%as the target recedes into the past\. In that synthetic evaluation, incidental benchmark properties can create substantial spurious advantages; rerunning the identical grid on SQuAD reproduces the ceiling in all 25 cells, though there the dense retriever does beat the lexical one, so that particular null was a property of our generator\. Bounded\-memory evaluations should hold access fixed and report retention and selection separately\.

## 1Introduction

Long\-running agents can accumulate more observations than they can keep available at once, so some must be dropped\. This makes memory a control problem: given a budget ofKKitems, which observations remain available later? Existing memory architectures each embed an assumption about what is worth keeping: that recent, repeated, or query\-resembling things matter\.

Evaluating those assumptions is harder than it looks, because an agent makes two decisions\. Aretentiondecision happens online, as each observation arrives, and is necessarily*query\-independent*: the future query has not been seen\. Aselectiondecision happens at query time and*can*be query\-aware\. A method that appears to win may be winning at either stage, or may simply have been granted access to material its competitors discarded\. One accuracy number per method cannot distinguish these\.

This is easy to introduce accidentally\. Auditing an earlier version of this evaluation, which compared five memory heuristics, we found that its strongest methods re\-ranked the complete history at query time while its weakest saw only a bounded buffer; that a template defect left its queries with no token identifying the fact asked about; and that its distractors carried a literal marker any overlap scorer could exploit\. Two of its five conditions in fact retained the entire history and so shared an access regime, yet were reported far apart, a gap that cannot reflect memory\-budget behaviour at all\. Each of these is mundane on its own\. Together they produced a large, clean, misleading result\.

We therefore rebuild the evaluation as a factorial design\. Our contributions:

- •A streaming\-recall benchmark that crossesretention rule×\\timesselection rule, evaluates every cell on the same seeded episodes, and reports recall alongside the retention rate that upper\-bounds it \(§[3](https://arxiv.org/html/2610.00366#S3)\)\.
- •A decomposition showing that most of the apparent advantage of query\-aware memory in the mixed\-cell comparison is attributable to history access rather than query awareness, with the remainder capped by retention \(§[4\.1](https://arxiv.org/html/2610.00366#S4.SS1)\)\.
- •Evidence that query\-aware retrieval gives no protection against delay\-induced eviction; that the retention ceiling is selection\-independent, verified against a dense retriever rather than assumed; and that incidental benchmark properties can create spurious apparent gains, with the stronger retriever amplifying one such artifact \(§[4\.2](https://arxiv.org/html/2610.00366#S4.SS2), §[4\.3](https://arxiv.org/html/2610.00366#S4.SS3)\); and that the ceiling result reproduces on real SQuAD text while the dense\-versus\-lexical null does not \(§[5](https://arxiv.org/html/2610.00366#S5)\)\.

## 2Related work

Agent memory systems score stored items by a mixture of signals:[Park et al\. \(2023\)](https://arxiv.org/html/2610.00366#bib.bib5)retrieve from a memory stream by weighted recency, importance, and relevance;[Packer et al\. \(2023\)](https://arxiv.org/html/2610.00366#bib.bib4)page between a bounded context and external storage;[Zhong et al\. \(2024\)](https://arxiv.org/html/2610.00366#bib.bib8)add a forgetting schedule\. These systems couple several memory\-management decisions within end\-to\-end evaluations, which is exactly the regime in which retention and retrieval effects become hard to separate\.

The selection half is classical retrieval, where lexical scorers such as BM25\([Robertson and Zaragoza, 2009](https://arxiv.org/html/2610.00366#bib.bib7)\)remain strong against dense retrievers\([Karpukhin et al\., 2020](https://arxiv.org/html/2610.00366#bib.bib1)\)and RAG\([Lewis et al\., 2020](https://arxiv.org/html/2610.00366#bib.bib2)\)conditions generation on retrieved evidence\. But that literature assumes a fixed corpus, with nothing thrown away; bounded memory breaks that assumption\. Relatedly,[Liu et al\. \(2024\)](https://arxiv.org/html/2610.00366#bib.bib3)show that where evidence sits in a context changes whether a model uses it: another case in which an apparent capability difference turns out to be a property of what was made available to the model\. We contribute to the evaluation layer rather than proposing a memory architecture; our object of study is what a bounded\-memory benchmark measures\.

## 3A factorial streaming\-recall benchmark

#### Environment\.

An episode is a stream of 10 facts followed by 2 queries\. Every fact has the identical surface form"Theslotisvalue\."; slots are drawn without replacement within an episode, so a query naming a slot identifies exactly one fact\. A fractiondd\(default0\.30\.3\) of facts are distractors, drawn from a disjoint slot vocabulary but written in the same form, so distractors are*not*lexically marked\. Each query targets a uniformly sampled non\-distractor fact\.

#### The two factors\.

Aretention ruleconsumes the stream online under capacityKKand is never shown the query:recency\(keep the lastKK\),random\(admit withp=0\.5p\{=\}0\.5, FIFO evict\),frequency\(evict the item whose tokens are rarest so far, FIFO on ties\),type\-oracle\(admit only non\-distractors; this reads the generator’s ground\-truth label and is therefore*privileged*, not implementable by an agent\), andunbounded\(keep everything; the no\-forgetting reference\)\. Aselection rulethen ranks the survivors and exposes the topk=3k\{=\}3:recency,random,query\(Jaccard overlap between query and fact tokens, recency tie\-break\),dense\(cosine similarity between MPNet sentence embeddings111all\-mpnet\-base\-v2, 768\-d\. The vocabularies are closed, so we cache all 353 realizable strings once; selection is then a dot product, and exactly deterministic\.\), andoracle\(surface the target if present\)\. We call the Jaccard rule*query\-similarity*and never an embedding method;denseis the only rule here that uses learned vectors\.

#### Query informativeness\.

We also vary whether the query carries usable signal\. An*informative*query is"What is theslot?"naming the target’s slot; an*uninformative*query is"What is the value?", wherevalueappears in no fact\. The two differ by exactly one token, isolating query informativeness from all else\.

#### Metric and statistics\.

We reportrequired\-fact recall: the fraction of queries whose target fact appears among thekkexposed items\. We report recall rather than answer accuracy because it isolates the memory pipeline; §[4\.4](https://arxiv.org/html/2610.00366#S4.SS4)then puts a language model at the end and shows recall predicts answer correctness up to a near\-constant factor\. Alongside recall we report theretention rate, the fraction of queries whose target survived retention, which upper\-bounds recall for any selection rule\. All cells use the same 300 seeded episodes \(600 queries\); confidence intervals are percentile bootstrap over 2,000 resamples*of episodes*, since two queries from one stream are not independent\. Seed, configuration, and per\-query records are saved with the results\.

## 4Results

### 4\.1RQ1: how much of the advantage is query awareness?

Table[1](https://arxiv.org/html/2610.00366#S4.T1)gives the full grid under informative queries\. Reading it by column isolates selection; by row, retention\.

Table 1:Required\-fact recall \(%\) under*informative*queries,K=5K\{=\}5,k=3k\{=\}3, 600 queries per cell on the same 300 episodes\. “Retention” is the fraction of queries whose target survived retention and upper\-bounds every entry in its row\.†type\-oraclereads the generator’s labels and is privileged, not implementable by an agent\. The 100% entries are a construction ceiling, not a performance result: an informative query token\-matches exactly one fact, so selection is exactly solvable when nothing was discarded\.Two things follow\. First, in every bounded row thequery,denseandoraclecolumns are*exactly equal*to the retention rate \(46\.846\.8,46\.846\.8,45\.845\.8,70\.370\.3\): selection saturates*in this environment*, and every remaining failure is a retention failure\. We test that with a far stronger scorer rather than asserting it: a 768\-dimensional MPNet retriever changes recall by\+0\.00\+0\.00pp\[\+0\.00,\+0\.00\]\[\+0\.00,\+0\.00\]against both the lexical rule and an oracle selector, under bounded and unbounded retention alike, and across all 50 cells no selection rule exceeded its row’s retention rate\. Saturation is a property of this generator, not a general one: §[5](https://arxiv.org/html/2610.00366#S5)shows that on real text a dense retriever does improve selection, and yet still cannot pass the retention ceiling\. The ceiling is the general claim; saturation is not\. A better retriever is not merely unhelpful; it is arithmetically incapable of helping, the exposed set being a subset of what retention kept\.

Second, a mixed\-cell comparison does not report the contrast it appears to\.unbounded\+queryagainstrecency\+recency, structurally what our earlier evaluation did, gives\+68\.7\+68\.7pp\[\+65\.2,\+72\.0\]\[\+65\.2,\+72\.0\]; fixing access at a bounded buffer and varying only selection gives\+15\.5\+15\.5pp\[\+12\.8,\+18\.2\]\[\+12\.8,\+18\.2\]; fixing selection and varying only access gives\+53\.2\+53\.2pp\[\+49\.7,\+56\.8\]\[\+49\.7,\+56\.8\]\. Three\-quarters of the headline effect is the freedom to re\-rank material the baselines had discarded \(Fig\.[1](https://arxiv.org/html/2610.00366#S4.F1)c\)\.

![Refer to caption](https://arxiv.org/html/2610.00366v1/fig_main.png)Figure 1:Required\-fact recall under query\-aware selection; the lexical and dense scorers are identical throughout under informative queries, so one set of curves serves for both\.\(a\)vs\. memory budgetKK; selection is inert atK≤k=3K\\leq k\{=\}3\.\(b\)vs\. delay \(facts observed after the target;n=188/159/253n\{=\}188/159/253\): recency retention falls to*zero*: query\-aware retrieval offers no protection against eviction\.\(c\)The apparent advantage atK=5K\{=\}5: adding query\-aware selection at matched access gives\+15\.5\+15\.5pp, adding full\-history access a further\+53\.2\+53\.2pp\.
### 4\.2RQ2: behaviour under memory pressure

Table 2:Recall \(%\) under query\-aware selection across memory budgetsKK, informative queries,k=3k\{=\}3\. AtK≤kK\\leq knothing retained is withheld, so selection is inert and every rule ties\.#### Budget\.

Recall under recency retention with query\-aware selection rises22\.7→31\.3→46\.8→100→10022\.7\\to 31\.3\\to 46\.8\\to 100\\to 100acrossK∈\{2,3,5,10,20\}K\\in\\\{2,3,5,10,20\\\}\(Table[2](https://arxiv.org/html/2610.00366#S4.T2), Fig\.[1](https://arxiv.org/html/2610.00366#S4.F1)a\)\. BelowK≤kK\\leq kthe selection rule is inert, because everything retained is already exposed, so the entire spread across selection rules at small budgets is zero\. This deserves stating plainly:*a query\-aware retriever can only matter when the buffer is larger than the retrieval budget\.*Every bounded rule converges to the unbounded ceiling onceKKreaches the stream length, which is the trivial regime in which forgetting has stopped happening at all\.

#### Delay\.

Bucketing queries by how many facts arrived after the target \(Fig\.[1](https://arxiv.org/html/2610.00366#S4.F1)b\), recency retention yields100%→58\.5%→0\.0%100\\%\\to 58\.5\\%\\to 0\.0\\%across delays of00–22,33–55, and66–99facts, and type\-oracle retention100%→97\.5%→31\.2%100\\%\\to 97\.5\\%\\to 31\.2\\%; unbounded retention is flat at100%100\\%\(Table[3](https://arxiv.org/html/2610.00366#S4.T3)\)\. The collapse to exactly zero is the sharpest result we have\. Selection is oracle\-equivalent in every one of these rows, so this is not a retrieval failure at all: by the time the query arrives the fact is gone, and no scorer, however strong, can rank an item that is no longer there\. Filtering distractors \(type\-oracle\) delays the cliff but does not remove it, falling to31\.2%31\.2\\%at the longest delay\.

Table 3:Recall \(%\) under query\-aware selection by delay, measured as the number of facts observed after the target,K=5K\{=\}5, informative queries \(n=188/159/253n=188/159/253queries per bucket\)\. Selection is oracle\-equivalent in every row, so the decline is retention loss, not retrieval failure\.Table 4:Recall \(%\) under query\-aware selection as distractor densityddrises,K=5K\{=\}5, informative queries\. Higher density*benefits*the privileged type\-oracle, since more of the stream is filtered before capacity is consumed, while recency retention is roughly unchanged\.
#### Distractor density\.

Raising density from0\.10\.1to0\.70\.7leaves recency retention flat \(47\.3→50\.547\.3\\to 50\.5\) but lifts type\-oracle retention from54\.054\.0to100\.0100\.0\(Table[4](https://arxiv.org/html/2610.00366#S4.T4)\), because discarding a larger share of the stream leaves proportionally more room for targets\. Raising density therefore*benefits*the privileged type\-oracle, because more of the stream can be filtered before capacity is consumed, while recency retention is left roughly unchanged\. The rule that gains is the one that needs the answer key to do it\.

### 4\.3RQ3: benchmark properties that create spurious apparent gains

Table 5:The same grid under*uninformative*queries, which name no fact\. Thequerycolumn collapses ontorecencyexactly: with no discriminative token every fact ties and the rule degenerates to its tie\-break\.densedoes not collapse, but gains nothing from the query either \(§[4\.3](https://arxiv.org/html/2610.00366#S4.SS3)\)\.Table 6:Distractor marking control,*uninformative*queries only\. Prefixing distractors with a literal marker string, as the original generator did, lifts the query\-aware rule by\+9\.8\+9\.8pp\[\+7\.7,\+12\.2\]\[\+7\.7,\+12\.2\]and\+9\.0\+9\.0pp\[\+7\.0,\+11\.3\]\[\+7\.0,\+11\.3\]while leaving every other rule unchanged\. The scorer is detecting the marker, not matching the query\.Under*uninformative*queries \(Table[5](https://arxiv.org/html/2610.00366#S4.T5)\) thequerycolumn collapses onto therecencycolumn exactly \(31\.331\.3,31\.331\.3,30\.830\.8,30\.230\.2,41\.241\.2down the rows\): with no discriminative token, all facts tie and the rule degenerates to its tie\-break\. Query\-aware selection contributes literally nothing, which is the regime our earlier evaluation was unknowingly in\.

Yet an apparent advantage can be restored without restoring any query signal, in two ways\. Re\-marking distractors with a literal prefix, as the original generator did, raises uninformative\-query recall for the query\-aware rule by\+9\.8\+9\.8pp\[\+7\.7,\+12\.2\]\[\+7\.7,\+12\.2\]unbounded and\+9\.0\+9\.0pp\[\+7\.0,\+11\.3\]\[\+7\.0,\+11\.3\]bounded, leaving every other rule unchanged \(Table[6](https://arxiv.org/html/2610.00366#S4.T6)\): the scorer is detecting the marker, not matching the query\.

The dense retriever reproduces this with no marker at all\. Under uninformative queries it does*not*collapse ontorecency, reaching42\.0%42\.0\\%against a31\.5%31\.5\\%chance level, readable as embeddings extracting signal where lexical overlap cannot\. They do not:0\.0%0\.0\\%of the items it exposes are distractors, against a30%30\\%base rate\. It filters distractors from the benchmark’s latent type structure, then picks at random among the77relevant facts:3/7=42\.9%3/7=42\.9\\%predicted,42\.0%42\.0\\%observed\. The stronger retriever amplifies the artifact rather than resisting it\.

### 4\.4RQ4: does recall predict answer correctness?

Recall is a retrieval\-side proxy\. To test whether it tracks task success we place FLAN\-T5\-large at the end of the pipeline \(greedy decoding, normalized exact match, 300 queries from the same episode set\) and score its answers \(Table[7](https://arxiv.org/html/2610.00366#S4.T7)\)\.

Table 7:Answer accuracy \(%\) with FLAN\-T5\-large at the end of the pipeline, greedy decoding, 300 queries per condition drawn from the same seeded episodes, normalized exact match\. “Acc\.∣\\midpresent” conditions on the required fact having reached the model and is close to constant across conditions, which is what licenses recall as the primary metric\.With no memory the model scores0\.0%0\.0\\%, so the task genuinely requires memory\. A bounded agent \(K=5K\{=\}5, query selection\) recalls the target for47\.3%47\.3\\%of queries and answers36\.3%36\.3\\%correctly; full context scores73\.7%73\.7\\%and the required fact alone74\.7%74\.7\\%, so nine irrelevant facts cost1\.01\.0pp and the bottleneck is memory, not context noise\. Accuracy conditioned on the fact having reached the model spans only70\.770\.7–74\.7%74\.7\\%, so under these tested conditions answer accuracy is recall scaled by a roughly constant factor, which is what licenses recall as the primary metric\. The bounded agent’s63\.7%63\.7\\%failure rate splits into52\.752\.7pp where the fact never arrived and12\.712\.7pp where it arrived and the model erred anyway: five in six failures are memory failures\.

### 4\.5RQ5: what explains the remaining failures?

We classify every failure of the bounded agent \(recency retention, query selection\) into memory\-specific categories: eviction, where the target never survived retention; displacement, where it was retained but pushed out of the top\-kk; and similarity confusion, where a distractor outscores the target under the query scorer itself\. All 319 failures are evictions\. Displacement and similarity confusion account for zero\. Swapping in a query\-independent selector on the same streams adds 93 displacement failures, so query awareness removes the selection\-stage failure mode outright while leaving the dominant one untouched\. With the language model attached, of 300 queries the target was absent 158 times and present\-but\-answered\-wrongly 38 times; the latter is the only failure here that is not about memory, and it matches the12\.712\.7pp answering term of §[4\.4](https://arxiv.org/html/2610.00366#S4.SS4)\.

## 5Does this survive real text?

Everything so far runs on a generator we wrote, which is the obvious objection\. We therefore rerun the identical grid on SQuAD v1\.1 dev\([Rajpurkar et al\., 2016](https://arxiv.org/html/2610.00366#bib.bib6)\), changing only where episodes come from\. For each of 300 questions we split the target paragraph into sentences, mark the sentence containing the answer span as the target, and pad the stream to ten sentences with sentences drawn from unrelated paragraphs\. The distractors are ordinary English on other topics, so nothing marks them, and the query is the real SQuAD question rather than a template\. The retention rules, selection rules, metric and bootstrap are imported and called unmodified\.

Table 8:The same grid on*real SQuAD text*: recall \(%\) of the answer\-bearing sentence, 300 questions from SQuAD v1\.1 dev,K=5K\{=\}5,k=3k\{=\}3\. Each stream mixes the target paragraph’s sentences with sentences from unrelated paragraphs, so nothing marks the distractors\. The retention ceiling holds in all 25 cells\. Unlike the synthetic benchmarkdensenow beatsquery, yet under a bounded buffer it still only reaches the ceiling, never past it\.The central claim survives, and one secondary claim does not\. The retention ceiling holds in all 25 cells: no selection rule anywhere exceeds its row’s retention rate \(Table[8](https://arxiv.org/html/2610.00366#S5.T8)\)\. Under a bounded recency buffer the dense retriever again lands exactly on the ceiling,\+0\.00\+0\.00pp\[\+0\.00,\+0\.00\]\[\+0\.00,\+0\.00\]against an oracle selector, and unbounded access is worth\+44\.3\+44\.3pp\[\+38\.3,\+50\.3\]\[\+38\.3,\+50\.3\]to it\. Query awareness at matched access is worth\+20\.3\+20\.3pp\[\+15\.7,\+25\.0\]\[\+15\.7,\+25\.0\], more than the\+15\.5\+15\.5pp we measured synthetically, which is what one would expect once queries stop being templates\.

What does*not*transfer is the dense\-versus\-lexical null\. On real text the dense retriever genuinely beats Jaccard, by\+6\.0\+6\.0pp\[\+3\.0,\+9\.3\]\[\+3\.0,\+9\.3\]with unbounded retention and\+1\.7\+1\.7pp\[\+0\.3,\+3\.3\]\[\+0\.3,\+3\.3\]under a bounded buffer\. Our synthetic finding that the two are exactly equivalent was therefore a property of that generator, where an informative query token\-matches its target exactly, and not a general fact\. The dense\-versus\-lexical equivalence observed synthetically therefore does not generalize to natural text: a better retriever does help at the selection stage there\. It still cannot exceed what retention kept, which is the claim the paper actually makes\.

## 6Limitations

The main environment is synthetic, single\-episode and small \(10 facts, 2 queries\), built so selection is exactly solvable under informative queries, hence the 100% ceilings; §[5](https://arxiv.org/html/2610.00366#S5)rebuts the strongest form of that objection but both environments are still single\-turn recall rather than multi\-session agent memory\. That isolates retention, but means our numbers*bound*what selection contributes rather than estimating it on a realistic corpus; with paraphrase, partial overlap, or multi\-hop queries the matched\-access query\-awareness effect would likely be larger than the\+15\.5\+15\.5pp we measure\. The main grid is retrieval\-side, and the answer\-accuracy check of §[4\.4](https://arxiv.org/html/2610.00366#S4.SS4)uses a single small model with one prompt and greedy decoding, so it establishes that recall tracks answer correctness in this setting rather than characterizing how any particular agent would behave\. Our “delay” is positional distance within a stream, not elapsed time or cross\-session persistence: this is bounded streaming memory, not long\-term memory\. A cross\-encoder reranker could differ from our two scorers, though the ceiling argument applies to any selector\. We excluded rather than reported an untrained linear write\-gate in the repository: it has no held\-out split and its reward comes from the answer key\.

## 7Conclusion

Retention and selection are different decisions, and one accuracy number cannot tell them apart\. Hold access fixed across compared methods, and report the retention rate beside every recall number\. When the required memory has been evicted a better retriever cannot recover it: retrieval performance is upper\-bounded by retention\.

## References

- Karpukhin et al\. \[2020\]V\. Karpukhin, B\. Oğuz, S\. Min, P\. Lewis, L\. Wu, S\. Edunov, D\. Chen, and W\. Yih\.Dense passage retrieval for open\-domain question answering\.In*EMNLP*, 2020\.
- Lewis et al\. \[2020\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela\.Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.In*NeurIPS*, 2020\.
- Liu et al\. \[2024\]N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. Liang\.Lost in the middle: How language models use long contexts\.*Transactions of the Association for Computational Linguistics*, 12:157–173, 2024\.
- Packer et al\. \[2023\]C\. Packer, S\. Wooders, K\. Lin, V\. Fang, S\. G\. Patil, I\. Stoica, and J\. E\. Gonzalez\.MemGPT: Towards LLMs as operating systems\.*arXiv preprint arXiv:2310\.08560*, 2023\.
- Park et al\. \[2023\]J\. S\. Park, J\. C\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein\.Generative agents: Interactive simulacra of human behavior\.In*UIST*, 2023\.
- Rajpurkar et al\. \[2016\]P\. Rajpurkar, J\. Zhang, K\. Lopyrev, and P\. Liang\.SQuAD: 100,000\+ questions for machine comprehension of text\.In*EMNLP*, 2016\.
- Robertson and Zaragoza \[2009\]S\. Robertson and H\. Zaragoza\.The probabilistic relevance framework: BM25 and beyond\.*Foundations and Trends in Information Retrieval*, 3\(4\):333–389, 2009\.
- Zhong et al\. \[2024\]W\. Zhong, L\. Guo, Q\. Gao, H\. Ye, and Y\. Wang\.MemoryBank: Enhancing large language models with long\-term memory\.In*AAAI*, 2024\.