EMIR$^2$: Evolution-Aware Memory with Intent-Guided Multi-Round Retrieval

• Long-term agents must maintain coherent memory as interactions update it and integrate relevant information scattered across historical events.Evolving information can become inconsistent or conflicting, while complex tasks may require evidence from multiple times and events.Source: §1 · p. 1
• EMIR2 jointly models factual evolution and adaptive evidence-seeking retrieval to address evolving memory and scattered historical evidence in long-horizon tasks.The framework combines SEMG’s evolving states, temporal event trajectories, and evidential associations with retrieval that expands based on unresolved information needs.Source: §1 · p. 2
• SEMG organizes episodic events into temporal chains and updates evolving semantic states with evidence, preserving factual trajectories for conflict-aware retrieval.This supports tracing how and why facts change, including replacement by newer facts, historical contradictions, and context-dependent preference shifts.Source: §1 · pp. 1–2, Figure 1 · p. 2
• Intent-guided multi-round retrieval identifies missing evidence and expands the search using accumulated evidence to support complex queries.This addresses single-round retrieval’s limited use of relationships among memories and its difficulty finding complementary evidence not directly matched to the original query.Source: §1 · pp. 2–3
• Experiments on LoCoMo and MemConflict demonstrate EMIR2’s effectiveness on complex questions and robustness to dynamic and static memory conflicts.Source: §1 · p. 3

2.1 MEMORY ORGANIZATION

• Memory systems evolved from episodic records toward workflows, interconnected networks, hierarchical tiers, and online/offline consolidation, but rarely explicitly model how related events evolve over time.MemoryBank adds forgetting and updating; Agent Workflow Memory reuses procedural experience; A-MEM organizes interconnected cards; MemoryOS migrates information across tiers; LightMem separates online processing from offline consolidation. SEMG organizes evolving events into temporal episodic chains with semantic nodes and optional lateral links.Source: §2.1 · p. 3

2.2 MEMORY RETRIEVAL

• Retrieval methods progress from similarity-based recall to structure-aware and adaptive strategies, yet rarely use memory-structure relationships to guide evidence exploration.LD-Agent (Li et al., 2025b) uses topic-based retrieval; SeCom uses segment-level memories and denoising; MemInsight adds semantic context; Mnemis combines similarity recall with semantic traversal; MRAgent iteratively explores and prunes paths. EMIR^2 retrieves over a chain-like memory graph, using temporal transitions and semantic relations to progressively discover relevant memories.Source: §2.2 · p. 3

3 PROBLEM DEFINITION

• The task is to build persistent memory from historical conversations and retrieve relevant evidence for a given query.Memory units preserve conversation content and contextual information.Source: §3 · p. 4
• Formally, the system constructs structured memory M = Build(D) from conversation history D = {d1, . . . , dN}.Source: §3 · p. 4
• For query q, retrieval identifies a compact set R = Retrieve(q, M), R ⊆ M, which is provided to the downstream model for answer generation.Source: §3 · p. 4

4.1 STATE-EVOLVING MEMORY GRAPH

• SEMG combines traceable episodic event chains with evidence-grounded semantic states to support retrieval and reasoning over evolving long-term interactions.Episodic memory preserves historical events, while semantic states track factual knowledge as it changes.Source: §4.1 · p. 4

4.1.1 EPISODIC MEMORY CONSTRUCTION

• Events extracted from local dialogue windows are assigned to topic-specific chains and ordered chronologically when temporal information is available.Temporal paths are complemented by classified relations—including causality, contextual relevance, and follow-ups—that connect events within or across topics.Source: §4.1.1 · p. 5

4.1.2 SEMANTIC STATE EVOLUTION

• Semantic states are updated by reinforcing or extending facts, revising replaced or retracted knowledge, or marking evidence insufficient for change.Revisions create successor states and preserve prior knowledge; event arrival alone does not require a new state.Source: §4.1.2 · p. 5

4.2 INTENT-GUIDED MULTI-ROUND RETRIEVAL

• Intent-guided retrieval iteratively assesses accumulated evidence, identifies gaps, and selects constrained graph operations until evidence is sufficient or the retrieval budget is exhausted.Retrieved nodes are deduplicated, kept within a total candidate budget, and reranked using model relevance and retrieval signals before the top K are returned.Source: §4.2 · p. 6

5.1 EXPERIMENT SETTINGS

• Evaluation covers LoCoMo’s single-hop, temporal, multi-hop, and open-domain questions, as well as overall accuracy, and uses MemConflict to assess handling of conflicting histories.LoCoMo evaluates retrieval from ultra-long dialogue memory; MemConflict evaluates handling of conflicting user histories.Source: §5.1 · p. 7

5.2 MAIN RESULTS

• EMIR2 shows particularly strong LoCoMo temporal and multi-hop results, remains competitive with a lightweight Qwen3.5-9B controller, and outperforms MRAgent overall under both backbone settings.With the Claude Sonnet 4.5 backbone, replacing the lightweight controller with the backbone LLM further improves aggregate multi-hop and temporal performance, but not uniformly across questions.Source: §5.2 · pp. 7–8

5.3 ABLATION STUDY

• Disabling multi-round retrieval or semantic nodes degrades performance, especially on multi-hop questions, while removing cross-chain edges causes no significant drop.The ablation variants test single-retrieval answering, removal of SEMG lateral edges, and removal of semantic nodes.Source: §5.3 · p. 8, Figure 3 · p. 8

5.4 CONFLICT-AWARE MEMORY EVALUATION

• On MemConflict, EMIR2 performs strongly in dynamic and static conflict scenarios, but conditional conflicts remain challenging because they require context-dependent filtering of valid information.The evaluation reports black-box answering accuracy and white-box retrieval performance.Source: §5.4 · pp. 8–9, Table 2 · p. 9

5.5 EFFECTIVENESS–EFFICIENCY TRADE-OFF OF MULTI-ROUND RETRIEVAL

• Increasing the retrieval-round budget initially improves evidence recall and answer accuracy, but gains become marginal beyond a threshold as token consumption rises.Adaptive termination lets EMIR2 stop when sufficient evidence has been collected, balancing evidence acquisition against retrieval overhead.Source: §5.5 · p. 9, Figure 4 · p. 9

6 CONCLUSION

• The framework combines temporal event chains, semantic states, and adaptive multi-round retrieval; experiments on LoCoMo and MemConflict show improved long-term reasoning and memory use under dynamic and static conflicts.Future work will explore finer-grained context-dependent memory modeling and efficient retrieval for scalable long-term agent systems.Source: §6 · p. 10

A.1 MEMORY RELATION TAXONOMY

• SEMG relations encode graph organization and event and state dependencies, including evidence links, causal connections, state support, and state revision or conflict.The taxonomy supports relation-aware graph traversal and retrieval.Source: Table 3 · p. 13, A MEMORY RELATION TAXONOMY AND… · p. 13, §A.1 · p. 14

A.2.1 WINDOW-LEVEL MEMORY EXTRACTION

• Memory construction uses five stages: window planning and extraction, entity normalization, semantic refinement, and cross-window event-relation determination.The extraction stage turns each historical window into episodic events, candidate semantic entries, and local relations; it separates temporary details into events and stable information into candidate claims.Source: §A.2 · p. 14, §A.2.1 · pp. 14, 16

A.2.2 SEMANTIC REDUCTION

• Semantic reduction reconciles incoming claims with the topic chain’s confirmed fact snapshot through reinforce, add, replace, or retract operations, leaving conflicting or insufficiently supported updates pending.An uncertain update retains its proposed operation without changing the active canonical state.Source: §A.2.2 · p. 17

A.2.3 CROSS-WINDOW EVENT-RELATION ADJUDICATION

• Cross-window relation adjudication independently checks candidate event pairs and relation direction against event summaries and evidence, conservatively rejecting unsupported relations.Candidate-generation signals are non-binding: the adjudicator may accept, override, or reject a proposed relation.Source: §A.2.3 · p. 18

B ADAPTIVE RETRIEVAL PROMPTS

• Adaptive retrieval uses model-based control for question decomposition, iterative evidence evaluation, graph expansion, semantic state refinement, and final evidence reranking.The appendix provides prompts for query decomposition, iterative retrieval control, graph-expansion source selection, and final reranking.Source: B ADAPTIVE RETRIEVAL PROMPTS · p. 18

B.1 QUESTION DECOMPOSITION GATE

• Before multi-round retrieval, a question-decomposition gate determines whether the query needs to be split into multiple retrieval targets.When decomposition is needed, the controller generates complementary sub-questions using the query, question type, retrieval requirements, and evidence from initial retrieval.Source: §B.1 · p. 19, §B.2 · p. 19

B.2 INITIAL SUBPROBLEM PLANNING

• Initial subproblem planning uses the detected question type, required information, and pre-retrieved evidence—including temporal details, salient facts, and retrieval provenance—to generate complementary sub-questions.This planning stage runs only after decomposition has been determined to be necessary.Source: §B.2 · p. 19
• Initial planning generates complementary subproblems that search for distinct facts, constraints, patterns, or explanations, and divides multi-event retrieval by event type or time range.Only information necessary to answer the root question is marked required.Source: §B.2 · p. 20

B.3 ITERATIVE EVIDENCE ASSESSMENT

• Evidence assessment checks whether the root question and required subproblems are supported, then derives focused retrieval probes and graph actions from any remaining gaps.Indirect evidence may jointly support inference, while factual questions require matches to the requested constraints; retrieval stops when the answer and required reasoning links are supported.Source: §B.3 · pp. 20–21

B.4 EXPANSION SOURCE SELECTION

• When evidence is insufficient, the controller selects up to five candidate memories whose graph neighborhoods are likely to expose missing evidence, even if the memories do not directly answer the question.It favors bridge, connector, entity, relation, causal, or temporal anchors, using candidate summaries and graph-edge metadata.Source: §B.4 · pp. 21–22

B.5 FINAL EVIDENCE RERANKING

• The final reranker independently scores every candidate for support of the original question, favoring direct or necessary indirect evidence over merely topical or redundant matches.For factual questions, scores depend on matching the requested subject, relation, facet, occurrence, ordinal, and temporal constraints; auxiliary modules also provide retrieval hints and slim semantic-state nodes to retain relevant facts and source evidence.Source: §B.5 · pp. 22–23

• The main experiments use Qwen3.5-9B as a lightweight retrieval controller, which slightly underperforms MRAgent on the multi-hop and temporal subsets.The capacity comparison replaces it with Claude Sonnet 4.5 and examines instance-level retrieval behavior to identify decisions that benefit from greater capacity.Source: C ANALYSIS OF RETRIEVAL CONTROLLER… · p. 23
• With Claude Sonnet 4.5 as retrieval controller, multi-hop score rises from 88.53 to 91.49 and temporal score from 85.25 to 87.54, exceeding corresponding MRAgent results.All other EMIR2 configurations remain unchanged; aggregate gains do not mean the stronger controller is better for every question.Source: Table 4 · p. 23, C ANALYSIS OF RETRIEVAL CONTROLLER… · p. 23

C.1 INSTANCE-LEVEL ANALYSIS

• The analysis characterizes retrieval behavior using one Qwen3.5-9B run with complete trajectories aligned to Claude runs by question, while Table 4 reports Qwen averages across three runs.The authors interpret the instance-level findings as improved retrieval decision quality in fine-grained scenarios, not a uniform benefit from more extensive retrieval.Source: §C.1 · pp. 23–24
• 31 predictions change from incorrect to correct with the Claude controller, while 16 change from correct to incorrect, so gains are concentrated in a subset of questions.The transitions compare Claude Sonnet 4.5 with Qwen3.5-9B.Source: Table 5 · p. 23, §C.1 · p. 23
• Claude uses fewer retrieval rounds on average while retaining complete supporting evidence for a larger proportion of questions and decomposing questions more often.Complete Gold is the proportion of questions for which all annotated supporting evidence is retained.Source: §C.1 · p. 23, Table 6 · p. 24
• More frequent decomposition helps acquire complementary evidence when questions involve multiple entities, events, or temporal constraints.The stronger controller more often turns the original question into more specific subquestions.Source: §C.1 · pp. 23–24

D ANALYSIS OF CROSS-CHAIN EDGES

• Cross-chain edges connect semantically related memories across topics, adding non-local reasoning paths that temporal chains alone cannot represent.Their contribution is to provide additional, more direct connections among related memories.Source: D ANALYSIS OF CROSS-CHAIN EDGES · p. 24
• Removing cross-chain edges reduces graph expansion and retrieved candidates, but final performance remains relatively stable because multi-round retrieval can explore alternative evidence paths.Without explicit cross-topic connections, the retriever can still find relevant memories through other relations and iterative retrieval.Source: Table 7 · p. 24, D ANALYSIS OF CROSS-CHAIN EDGES · p. 24
• Cross-chain edges contribute additional evidence paths, while most evidence remains recoverable through other retrieval processes and adaptive iterative exploration.This compensatory retrieval moderates the effect of missing connections on final answer quality.Source: Table 8 · p. 24, D ANALYSIS OF CROSS-CHAIN EDGES · p. 24

• MemConflict evaluation covers black-box and white-box results; all metrics are higher-is-better, and average scores span the three conflict types.Source: E DETAILED MEMCONFLICT EVALUATION · p. 25

• Black-box performance is assessed with AA for answer correctness, UOCS for following temporal update order in dynamic conflicts, and CRS for resolving static contradictions.Source: §E.1 · p. 25
• Table 9 reports MemConflict black-box performance by conflict type, marking the best overall methods in bold and the best baselines with underlining.Source: Table 9 · p. 25

E.2 WHITE-BOX RETRIEVAL ANALYSIS

• SEH@3 measures supporting-evidence inclusion in the top three retrieved memories, while SRS measures its ranking quality; Table 10 reports both by MemConflict conflict type.The table caption states that bold values mark the best performance overall and underlined values mark the best among baselines.Source: §E.2 · p. 25, Table 10 · p. 25