Retrieval Is Not Enough

On day 100, Lia asks Aurora:

What should you suggest tonight?

A vector search returns two highly similar statements that Lia avoids caffeine at night. A keyword search finds that she prefers tea without sugar. A graph traversal reaches a previous episode in which mint tea worked well. Another result says that she drinks coffee in the morning. A suspiciously relevant candidate says that she likes espresso, although no source supports it.

Returning the first three rows is easy. Building useful evidence is not. Two rows repeat one restriction. The espresso claim is unsupported. The morning habit is not a contradiction because its context differs. Each accepted claim also needs time, scope, and provenance. All of that must fit inside a finite context budget.

Retrieval therefore produces candidates, not context. A read pipeline must interpret the task, generate candidates through complementary indexes, apply eligibility gates, rerank for decision utility, select a diverse subset, close its evidence paths, serialize it deliberately, and report omissions.

This is the seventh article in the Memory in Graphs series and the first in Part III. Memory for Agents argued that reading memory means reconstructing evidence. The Memory Multigraph gave claims, events, episodes, and sources independent identities. Time, Provenance, and Contradiction defined valid historical views. How Experience Enters Memory built the write path. We now build the read path.

1. A retrieved item is not an evidence packet

Let (q) be a query and (\mathcal M) the memory graph. A retriever returns a ranked candidate sequence:

[ R(q,\mathcal M)=(c_1,c_2,\ldots,c_n). ]

The rank states that an item matched a retrieval function. It does not state that the item is current, authorized, supported, nonredundant, or useful for the decision. Dense Passage Retrieval showed that learned dense representations can improve passage recall over sparse baselines. Retrieval augmented generation showed how retrieved nonparametric content can condition generation. Both begin with retrieval because a reader cannot inspect an entire collection. Neither makes rank equivalent to evidential sufficiency.

We need a different output:

[ P=(H,E,Q,O), ]

where (H) is a small header describing query scope, (E) is an ordered evidence set, (Q) contains qualifiers and uncertainty, and (O) reports important omissions. We call (P) an evidence packet.

The difference is operational. A candidate can be highly similar and still fail authorization. Two candidates can be individually relevant and jointly redundant. A claim can be correct but useless without its temporal scope. A graph node can be central but unsupported by any source. Retrieval finds possible ingredients. Packet construction decides which complete ingredients the model receives.

1.1 Solved exercises

Exercise 1.1. What does rank 1 prove?

Solution. Only that the candidate scored highest under the chosen retriever. It does not prove truth or eligibility.

Exercise 1.2. Why is a source link part of the packet?

Solution. It lets the model and auditor distinguish supported memory from an isolated assertion.

Exercise 1.3. Can two relevant items reduce packet quality?

Solution. Yes. If they repeat the same evidence, they consume budget without adding coverage.

Exercise 1.4. Which object is the final product of the read path?

Solution. An ordered evidence packet, not the raw ranked list.

Exercise 1.5. Why include (O)?

Solution. Omission reasons reveal whether evidence was invalid, unauthorized, redundant, or merely too expensive.

2. The query contract

“Tonight” is not a bag of tokens. It identifies a temporal scope. “Lia” identifies a subject. “Suggest” requests a decision. The query contract makes these commitments explicit:

[ Q=(a,\tau,\kappa,t_v,t_t,\alpha,B), ]

where (a) is the action, (\tau) is the target entity, (\kappa) is context, (t_v) is the modeled valid instant, (t_t) is the knowledge instant, (\alpha) is read authorization, and (B) is the context budget.

For the canonical query, (a) is beverage recommendation, (\tau) is Lia, (\kappa) is evening, (t_v=t_t=100), and (B=12) abstract units. These units can later become tokens or bytes. Abstract units keep the arithmetic visible: every complete evidence bundle costs four, so exactly three bundles fit.

Query interpretation may expand terms. “Tonight” can add evening and night. “Suggest” can seek preferences, constraints, prior recommendations, and outcomes. Expansion must remain traceable. Otherwise a generated search term silently changes the task.

A query contract also states what not to retrieve. Another user’s memory is outside authorization. A future valid interval is outside (t_v). A version recorded after (t_t) is outside the historical knowledge view. Search engines prefer recall. Contracts supply boundaries.

2.1 Solved exercises

Exercise 2.1. What does (t_v=100,t_t=90) ask?

Solution. It asks what the system knew on day 90 about modeled day 100.

Exercise 2.2. Why separate action from keywords?

Solution. Recommendation, explanation, correction, and deletion can use the same words while requiring different evidence.

Exercise 2.3. What does (B=12) permit in our example?

Solution. Three complete bundles at four units each.

Exercise 2.4. Is query expansion free?

Solution. No. It changes recall behavior and must be versioned and observable.

Exercise 2.5. Which contract field prevents cross user retrieval?

Solution. Read authorization (\alpha), together with tenant and subject scope.

3. Candidate generation and rank fusion

No single retrieval representation owns relevance. Sparse retrieval preserves exact terms. Dense retrieval captures paraphrases. Temporal indexes narrow intervals. Graph traversal follows identity, episode, provenance, and relation structure.

Let ranker (r_j) assign a position (\operatorname{rank}_j(c)). Reciprocal Rank Fusion combines rankings without requiring comparable raw scores:

[ \operatorname{RRF}(c)= \sum_j\frac{1}{k+\operatorname{rank}_j(c)}, ]

where (k>0) reduces the effect of extreme top ranks.

Fusion improves candidate generation because “caffeine free” and “avoids caffeine” can arrive through different channels. It still cannot decide validity. An obsolete claim can rank highly in every index. An unsupported statement can match perfectly. Fusion answers, “Which items should the next stage inspect?”

Graph search adds another distinction. A direct claim may be one hop from Lia. Its source and outcome episode may be several hops away. Candidate generation should recall both seeds and plausible connecting paths, but it should not flood the packet with every neighbor. Graph expansion belongs after query scoping and before final cost selection.

3.1 Solved exercises

Exercise 3.1. Why use RRF instead of adding raw scores?

Solution. Sparse, dense, temporal, and graph scores may have incomparable scales, while ranks share a simple ordering.

Exercise 3.2. For (k=60), what is the contribution of rank 1?

Solution. (1/(60+1)\approx0.01639).

Exercise 3.3. Does a high fused score pass temporal validation?

Solution. No. Fusion operates before eligibility.

Exercise 3.4. Which retriever best finds an exact product code?

Solution. A sparse or structured index usually preserves exact lexical identity better.

Exercise 3.5. Why include graph paths in the candidate pool?

Solution. Multi step decisions may need bridge entities, episodes, and sources that are not textually similar to the query.

4. Eligibility before relevance

Eligibility is a predicate, not a ranking feature. At valid day (t_v) and knowledge day (t_t), define:

[ \mathcal E= {c\in\mathcal C\mid A(c)\land V(c,t_v,t_t)\land S(c)\land L(c)}, ]

where (A) checks authorization, (V) checks bitemporal visibility, (S) requires evidential support, and (L) requires an active lifecycle state.

The espresso preference fails (S). The original claim that the night restriction began on day 52 fails (L) and its transaction interval at day 100. Both disappear before reranking. Calling these gates penalties would be dangerous. A sufficiently high similarity could compensate for a penalty. It cannot compensate for missing permission or impossible time.

Eligibility also protects cost. There is no reason to expand provenance, fetch source text, or run an expensive cross encoder for a candidate that policy forbids. Cheap hard gates should run early.

A gate failure remains observable. Repeated temporal failures may indicate a query clock bug. Repeated unsupported candidates may indicate a broken consolidation job. Exclusion protects the packet; reason codes protect operations.

4.1 Solved exercises

Exercise 4.1. May relevance compensate for missing authorization?

Solution. No. Authorization is a hard gate.

Exercise 4.2. Why gate before expensive reranking?

Solution. It avoids spending compute on candidates that cannot enter the packet.

Exercise 4.3. A claim is valid on day 100 but recorded on day 105. Is it visible at (t_t=100)?

Solution. No. Its transaction interval has not begun.

Exercise 4.4. What does (S(c)) require?

Solution. A valid support path to source evidence, not merely a populated confidence field.

Exercise 4.5. Why log gate reasons?

Solution. They distinguish healthy filtering from systematic indexing, policy, or temporal defects.

5. Reranking for decision utility

Candidate generation favors recall. Reranking estimates how much each eligible candidate helps the present decision. A simple normalized score is:

[ r(c)=w_s s(c)+w_t t(c)+w_u u(c)+w_g g(c), \qquad \sum w_i=1, ]

where (s) is semantic relevance, (t) temporal fit, (u) task utility, and (g) graph support.

This formula is pedagogical. Production systems may use a learned cross encoder, pairwise ranker, or rule model. The important separation is that query similarity and decision utility are not identical. The mint tea outcome has slightly lower similarity than a caffeine paraphrase, but it adds evidence about what previously worked.

Scores need calibration or at least stable interpretation within a version. A value of (0.9) from one model is not automatically comparable with (0.9) from another. Rank fusion avoids that problem during recall. Reranking reintroduces it and must document normalization.

Reranking remains itemwise unless it explicitly considers sets. It can place every caffeine paraphrase above the outcome. Diversity and bundle cost belong to selection.

5.1 Solved exercises

Exercise 5.1. Why can a lower similarity item have higher utility?

Solution. It may supply a complementary constraint, outcome, or qualifier needed for the decision.

Exercise 5.2. If weights sum to 1, what useful property follows?

Solution. A convex combination of unit interval features remains in the unit interval.

Exercise 5.3. Does reranking remove duplicates automatically?

Solution. No. An itemwise ranker can place several duplicates at the top.

Exercise 5.4. What does temporal fit measure?

Solution. It measures how well a valid candidate’s time and context align with the query, after basic validity has passed.

Exercise 5.5. Why version score normalization?

Solution. Model or feature changes can shift score distributions and silently change thresholds.

6. Redundancy, novelty, and diversity

Maximal Marginal Relevance, MMR, balances query relevance against similarity to the selected set:

[ \operatorname{MMR}(c\mid S)= \lambda\operatorname{Rel}(c,q) -(1-\lambda)\max_{s\in S}\operatorname{Sim}(c,s). ]

When (S) is empty, the strongest night constraint enters. Its paraphrase then pays a redundancy penalty. Tea preference and mint outcome add new evidence roles, so their marginal values remain high.

Redundancy is not only textual. Two claims derived from the same source root do not provide independent support. Two different phrases may express one proposition. Two identical sentences may describe separate episodes. A memory packet therefore compares semantic identity, evidence role, source root, episode identity, and temporal scope.

Diversity is also not decoration. It improves coverage only when dimensions correspond to the task. Selecting one item from every arbitrary cluster can displace crucial evidence. Our selector rewards new roles among constraint, preference, outcome, and qualifier because these roles have decision meaning.

6.1 Solved exercises

Exercise 6.1. What happens as (\lambda\to1)?

Solution. MMR approaches relevance only ranking.

Exercise 6.2. What happens as (\lambda\to0)?

Solution. Novelty dominates and may select diverse but weakly relevant items.

Exercise 6.3. Are two sources with equal text necessarily duplicates?

Solution. No. They may be independent occurrences, although proposition content can still be redundant for the packet.

Exercise 6.4. Why penalize a shared source root?

Solution. Multiple derivations from one origin should not masquerade as independent evidence.

Exercise 6.5. Which roles fill the canonical packet?

Solution. Constraint, preference, and outcome.

7. Graph expansion and evidence closure

A selected claim is not complete by itself. Let (\operatorname{cl}(c)) be the minimal graph closure required to interpret and audit it. Closure can contain subject identity, evidence source, extraction activity, temporal qualifier, episode, and relation occurrence.

For a set (S):

[ \operatorname{cl}(S)=\bigcup_{c\in S}\operatorname{cl}(c). ]

The marginal cost of adding candidate (c) is:

[ \Delta b(c\mid S)= \operatorname{cost}\left( \operatorname{cl}(c)\setminus\operatorname{cl}(S) \right). ]

Shared closure matters. Two claims from one source can reuse the same source header. A new episode may require several bridge nodes. Counting only claim text systematically underestimates graph evidence.

Closure rules must be typed and bounded. Following every provenance ancestor can reach model builds, deployments, and organizations. The query needs enough provenance to support use, while a full audit can request a deeper packet. Evidence profiles make this distinction explicit.

Joint evidence retrieval research shows why sets and chains matter for multi step questions. A set of individually strong sentences can omit the bridge that makes them jointly interpretable. Graph closure turns that bridge into a required cost rather than a lucky retrieval.

7.1 Solved exercises

Exercise 7.1. Why is claim text alone insufficient?

Solution. It can omit identity, scope, source, or the episode connecting the claim to the task.

Exercise 7.2. When can marginal closure cost be smaller than standalone cost?

Solution. When the packet already contains shared source, entity, or episode fragments.

Exercise 7.3. Why bound provenance traversal?

Solution. Full ancestry may be large and irrelevant to the immediate decision.

Exercise 7.4. What is an evidence profile?

Solution. A policy defining which closure types and depths a task requires.

Exercise 7.5. What does a bridge node contribute?

Solution. It connects individually retrieved facts into a coherent evidence chain.

8. Contradictions, qualifiers, and missing evidence

The morning coffee habit appears close to caffeine and Lia. It does not contradict a night restriction because context differs. A packet can include it as a qualifier when the task needs to prevent overgeneralization.

Contradiction candidates require aligned subject, proposition, polarity, modality, context, and overlapping valid time. When a genuine conflict exists, the selector should not keep only the higher scoring side. It should include both supported sides or report that the budget cannot represent the dispute safely.

Missing evidence also matters. Suppose the packet finds a positive outcome but cannot resolve which recommendation produced it. The honest serialization says “prior mint tea outcome, episode unresolved.” It does not complete the path through invention.

An omission report distinguishes:

• ineligible items removed by policy;
• redundant items displaced by stronger coverage;
• valid items that exceeded budget;
• required closure that could not be fetched;
• unresolved conflicts that prevent a confident answer.

Abstention can therefore be a retrieval result. If no packet satisfies the evidence profile, the read path should say so before generation.

8.1 Solved exercises

Exercise 8.1. Does morning coffee contradict no caffeine at night?

Solution. No. Their contexts are disjoint.

Exercise 8.2. What should happen to two genuinely conflicting supported claims?

Solution. Include both with scope and provenance, or abstain if the budget cannot represent the conflict safely.

Exercise 8.3. May retrieval fill an unresolved episode link?

Solution. No. It must preserve the unresolved state.

Exercise 8.4. Why report budget omissions?

Solution. The caller can distinguish absent memory from evidence that existed but did not fit.

Exercise 8.5. When should packet construction abstain?

Solution. When required evidence, authorization, closure, or conflict coverage cannot be satisfied.

9. Selection under a context budget

Let each closure fragment have cost. We seek:

[ \max_{S\subseteq\mathcal E}F(S) \quad\text{subject to}\quad \operatorname{cost}(\operatorname{cl}(S))\le B. ]

(F) rewards relevance, utility, role coverage, and novelty. This is a budgeted subset selection problem. Exact optimization may be expensive, but greedy marginal selection is transparent and often practical:

[ c^*= \arg\max_{c\notin S} \frac{\Delta F(c\mid S)}{\Delta b(c\mid S)}. ]

Our bundles all cost four, so value and value per cost produce the same order. At (B=12), the relevance baseline selects memory 1, its paraphrase memory 2, and unsupported memory 6. The gated diverse selector chooses memory 1, tea preference memory 3, and mint outcome memory 4.

The example is deliberately small. Real token cost depends on serialization, tokenizer, shared headers, and compression. Estimate after formatting or maintain a conservative bound. A packet that exceeds the model limit after adding labels was never budget compliant.

9.1 Solved exercises

Exercise 9.1. How many four unit bundles fit in (B=12)?

Solution. Three.

Exercise 9.2. Why divide marginal value by marginal cost?

Solution. It compares the evidence gained per remaining budget unit.

Exercise 9.3. When is standalone cost misleading?

Solution. When selected bundles share closure fragments or require different bridge costs.

Exercise 9.4. Why tokenize after serialization?

Solution. Labels, provenance fields, and formatting also consume tokens.

Exercise 9.5. What does the canonical selector omit after filling the budget?

Solution. The redundant constraint paraphrase and morning qualifier, with explicit budget reasons.

10. Ordering and serialization

Selection chooses a set. A model receives a sequence. Ordering is therefore part of the read algorithm.

The canonical packet leads with the hard night constraint, follows with the durable tea preference, and ends its decision evidence with the prior mint tea outcome. Each item includes time, scope, confidence, source identity, and a compact evidence excerpt. Qualifiers follow the proposition they qualify. Conflicting claims remain adjacent.

Long context experiments by Nelson Liu and coauthors found that model performance can change substantially with evidence position, often degrading when relevant information appears in the middle. We should not turn that result into a universal magic ordering. We should test packet templates with the actual model and task.

Stable ordering improves reproducibility and cache behavior. A deterministic tie break by evidence role, score, time, and identity prevents equivalent packets from changing merely because a hash table iterated differently.

The pipeline laboratory exposes all seven stages:

10.1 Solved exercises

Exercise 10.1. Why is a selected set not ready for a language model?

Solution. The model consumes an ordered serialization with labels and framing.

Exercise 10.2. Where should a qualifier appear?

Solution. Near the proposition it qualifies so their relationship survives serialization.

Exercise 10.3. Does one ordering work for every model?

Solution. No. Position sensitivity must be evaluated for the deployed model and task.

Exercise 10.4. Why use deterministic tie breaks?

Solution. They support reproducibility, caching, and regression tests.

Exercise 10.5. Which item leads the canonical packet?

Solution. The valid night caffeine constraint because it is a direct decision boundary.

11. Caching, staleness, and observability

Caching a packet can save retrieval and reranking cost, but a packet is valid only under its contract. A cache key must include query interpretation, memory version, valid and knowledge times, authorization scope, evidence profile, budget, retriever version, and serializer version.

The dangerous cache key is normalized query text alone. “What should you suggest tonight?” can require a different answer after a correction, permission change, or new outcome. Memory is designed to change. A cache that ignores change defeats the design with impressive efficiency.

Observe each stage separately: candidate counts by retriever, gate failures by reason, reranker distributions, redundancy removed, closure expansion, budget utilization, omission classes, serialization tokens, and downstream citations. These traces let us distinguish missing recall from poor selection.

Latency should also be decomposed. Candidate generation, graph fetch, reranking, closure, and serialization have different scaling behavior. Optimizing vector search while a cross encoder dominates total time is polishing the fastest part of the path.

11.1 Solved exercises

Exercise 11.1. Why include memory version in a cache key?

Solution. New claims, corrections, and deletions can change eligibility and selection.

Exercise 11.2. Which metric reveals excessive duplication?

Solution. Redundancy removed or unique evidence roles before and after selection.

Exercise 11.3. Why time each stage?

Solution. Total latency alone cannot identify the component worth optimizing.

Exercise 11.4. Can authorization changes invalidate a cached packet?

Solution. Yes. Read scope is part of the packet contract.

Exercise 11.5. What does low budget utilization indicate?

Solution. It may indicate insufficient eligible evidence, coarse bundle granularity, or a selector defect.

12. A C++23 reference implementation

The reference program is a deterministic CPU correctness baseline. MemoryUnit stores strong identities, kind, complete bundle cost, relevance, utility, two temporal intervals, and gate predicates. Checked construction uses std::expected.

The relevance baseline sorts pointers into contiguous storage and fills 12 units without validation. It selects:

memory 1 | constraint | Lia avoids caffeine at night.
memory 2 | constraint | Evening drinks for Lia should be caffeine free.
memory 6 | preference | Lia likes espresso.

The packet builder filters first, then recomputes marginal value after every selection:

const double coverage =
    already_has_kind(selected, candidate.kind) ? 0.0 : 0.50;
const double kind_redundancy =
    already_has_kind(selected, candidate.kind) ? 0.60 : 0.0;
const double source_redundancy =
    already_has_source(selected, candidate.source_root) ? 0.70 : 0.0;

Its output is:

memory 1 | constraint | Lia avoids caffeine at night.
memory 3 | preference | Lia prefers tea without sugar.
memory 4 | outcome    | Mint tea worked well in a prior episode.

The complete source is evidence_packet.cpp. It uses std::span for nonowning views, ranges algorithms for sorting and predicates, contiguous std::vector ownership, and assertions for every canonical invariant. It compiles in current MSVC with C++ latest mode, strict conformance, and warning level 4.

The CPU owns policy, graph closure, and serialization. Dense encoders or rerankers may run on GPUs, but accelerator scores return to this control path. GPU throughput does not authorize a claim, repair time, or explain an omission.

12.1 Solved exercises

Exercise 12.1. Why store pointers into one vector?

Solution. Ranking can reorder lightweight references while stable contiguous ownership keeps object lifetimes clear.

Exercise 12.2. Why recompute marginal value?

Solution. Novelty and shared source penalties depend on what the packet already contains.

Exercise 12.3. What prevents a superseded claim from entering the packet?

Solution. The active lifecycle gate and transaction time check.

Exercise 12.4. Why serialize after selection?

Solution. Selection optimizes membership; serialization imposes a stable task specific order.

Exercise 12.5. Is the fixed example a performance benchmark?

Solution. No. It validates semantics and invariants, not production latency or scale.

13. Stagewise evaluation, limits, and the frontier

Classical retrieval measures remain necessary. For a relevant set (G) and retrieved set (R_k):

[ \operatorname{Recall@}k=\frac{|R_k\cap G|}{|G|}, \qquad \operatorname{Precision@}k=\frac{|R_k\cap G|}{k}. ]

They are not sufficient. Evaluate eligibility precision, temporal accuracy, authorization violations, provenance coverage, evidence role coverage, redundancy, budget utilization, omission accuracy, packet faithfulness, downstream answer correctness, and abstention.

LongMemEval separates extraction, multi session reasoning, temporal reasoning, knowledge updates, and abstention. That separation matters because one answer score can hide whether failure began at writing, recall, selection, or reading. BERGEN likewise emphasizes reproducible evaluation across retrievers, rerankers, generators, datasets, and metrics.

Packet construction still has limits. Relevance labels can be incomplete. Utility is task dependent. Graph closure can grow explosively. Contradictions may not be known until reasoning begins. A language model can ignore perfect evidence or generate unsupported content. Retrieval is a necessary interface, not a guarantee of reasoning.

MG07 gives us a disciplined read path. MG08, Episodic Memory as an Event Graph, will now examine the objects that this path increasingly needs: ordered actions, observations, outcomes, and causal claims inside episodes that retain their situated history.

13.1 Solved exercises

Exercise 13.1. If three of four relevant items appear in the top five, what is Recall@5?

Solution. (3/4=0.75).

Exercise 13.2. What is Precision@5 in the same case?

Solution. (3/5=0.60).

Exercise 13.3. Can perfect recall produce a bad packet?

Solution. Yes. The packet can still be redundant, invalid, over budget, poorly ordered, or unsupported.

Exercise 13.4. Which measure catches missing sources?

Solution. Provenance coverage or support path completeness.

Exercise 13.5. What does MG08 add?

Solution. A complete event graph for episodes, including order, participants, actions, observations, outcomes, and cautious causal structure.

Conclusion

Retrieval is a recall mechanism. Memory reading is evidence construction.

We interpret the query before search. We combine retrieval channels without treating fusion as truth. We apply authorization, bitemporal, provenance, and lifecycle gates before expensive scoring. We rerank for decision utility. We penalize semantic and source redundancy. We expand selected claims to bounded graph closure. We choose bundles under their complete cost. We serialize deliberately and preserve omission reasons.

The result is smaller than the candidate pool and more useful than a top-k list. It can tell Aurora not only that a night caffeine restriction matched the query, but also that the restriction is current, supported, scoped to evening, distinct from the morning coffee habit, and compatible with a prior successful mint tea episode.

A memory has not been retrieved until its evidence can survive the decision it is meant to support.

References

• CORMACK, Gordon V.; CLARKE, Charles L. A.; BÜTTCHER, Stefan. Reciprocal rank fusion outperforms Condorcet and individual rank learning methods. In: Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2009. p. 758–759. DOI: 10.1145/1571941.1572114. Available at: https://cormack.uwaterloo.ca/cormacksigir09-rrf.pdf. Accessed on: 23 July 2026.
• GOLDSTEIN, Jade; CARBONELL, Jaime. The use of MMR, diversity based reranking for reordering documents and producing summaries. In: Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 1998. Available at: https://www.cs.cmu.edu/~jgc/publication/The_Use_MMR_Diversity_Based_LTMIR_1998.pdf. Accessed on: 23 July 2026.
• JIMÉNEZ GUTIÉRREZ, Bernal; SHU, Yiheng; GU, Yu; YASUNAGA, Michihiro; SU, Yu. HippoRAG: neurobiologically inspired long term memory for large language models. arXiv, 2024. DOI: 10.48550/arXiv.2405.14831. Available at: https://arxiv.org/abs/2405.14831. Accessed on: 23 July 2026.
• KARPUKHIN, Vladimir; OGUZ, Barlas; MIN, Sewon; LEWIS, Patrick; WU, Ledell; EDUNOV, Sergey; CHEN, Danqi; YIH, Wen-tau. Dense passage retrieval for open domain question answering. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Online: Association for Computational Linguistics, 2020. p. 6769–6781. DOI: 10.18653/v1/2020.emnlp-main.550. Available at: https://aclanthology.org/2020.emnlp-main.550/. Accessed on: 23 July 2026.
• LEWIS, Patrick; PEREZ, Ethan; PIKTUS, Aleksandra; PETRONI, Fabio; KARPUKHIN, Vladimir; GOYAL, Naman; KÜTTLER, Heinrich; LEWIS, Mike; YIH, Wen-tau; ROCKTÄSCHEL, Tim; RIEDEL, Sebastian; KIELA, Douwe. Retrieval augmented generation for knowledge intensive NLP tasks. In: Advances in Neural Information Processing Systems 33. [S. l.]: NeurIPS, 2020. Available at: https://papers.nips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html. Accessed on: 23 July 2026.
• LIU, Nelson F.; LIN, Kevin; HEWITT, John; PARANJAPE, Ashwin; BEVILACQUA, Michele; PETRONI, Fabio; LIANG, Percy. Lost in the middle: how language models use long contexts. Transactions of the Association for Computational Linguistics, v. 12, p. 157–173, 2024. DOI: 10.1162/tacl_a_00638. Available at: https://aclanthology.org/2024.tacl-1.9/. Accessed on: 23 July 2026.
• MANNING, Christopher D.; RAGHAVAN, Prabhakar; SCHÜTZE, Hinrich. Introduction to information retrieval. Cambridge: Cambridge University Press, 2008. Available at: https://nlp.stanford.edu/IR-book/pdf/irbookonlinereading.pdf. Accessed on: 23 July 2026.
• RAU, David; DÉJEAN, Hervé; CHIRKOVA, Nadezhda; FORMAL, Thibault; WANG, Shuai; CLINCHANT, Stéphane; NIKOULINA, Vassilina. BERGEN: a benchmarking library for retrieval augmented generation. In: Findings of the Association for Computational Linguistics: EMNLP 2024. Miami: Association for Computational Linguistics, 2024. p. 7640–7663. DOI: 10.18653/v1/2024.findings-emnlp.449. Available at: https://aclanthology.org/2024.findings-emnlp.449/. Accessed on: 23 July 2026.
• WU, Di; WANG, Hongwei; YU, Wenhao; ZHANG, Yuwei; CHANG, Kai-Wei; YU, Dong. LongMemEval: benchmarking chat assistants on long term interactive memory. arXiv, 2024. DOI: 10.48550/arXiv.2410.10813. Available at: https://arxiv.org/abs/2410.10813. Accessed on: 23 July 2026.
• YADAV, Vikas; BETHARD, Steven; SURDEANU, Mihai. If you want to go far go together: unsupervised joint candidate evidence retrieval for multi hop question answering. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics. Online: Association for Computational Linguistics, 2021. p. 4571–4581. DOI: 10.18653/v1/2021.naacl-main.363. Available at: https://aclanthology.org/2021.naacl-main.363/. Accessed on: 23 July 2026.