RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory

Jun Liu; Zhenglun Kong; Changdi Yang; Fan Yang; Tianqi Li; Peiyan Dong; Joannah Nanjekye; Hao Tang

cs.CLAug 6, 2025

RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory

Authors: Jun Liu, Zhenglun Kong, Changdi Yang, Fan Yang, Tianqi Li, Peiyan Dong, Joannah Nanjekye, Hao Tang,

Organizations: Carnegie Mellon University · Northeastern University · Harvard University · Fujitsu Research of America ·

Abstract

Large language model (LLM) agents are increasingly organized into multi-agent systems for complex reasoning and decision-making tasks. However, scaling such systems can substantially increase inference cost, especially when multiple agents repeatedly process overlapping or role-irrelevant context. Many existing multi-agent frameworks focus on orchestration and role specialization, but still rely on full-context sharing or static context routing, leading to redundant token consumption, increased runtime, and inefficient memory utilization. We present Role-Conditioned Context Routing (RCR), an execution-time layer that allocates role-relevant shared-memory items to agents under token budgets. Given a structured shared memory and a set of role-specialized agents, RCR selectively routes task-relevant memory items to each agent according to its role, optional task-stage metadata, and token budget. RCR integrates role-conditioned relevance scoring, token-budgeted memory selection, and iterative memory updates to reduce redundant context exposure during multi-agent execution. Experiments on multi-hop QA, embodied reasoning (ALFWorld), and web-based decision-making (WebShop) show that RCR reduces token usage by 35% on average and runtime by 38% on average relative to Full-Context, while maintaining or improving task performance.

Figures & tables

Figure 1: Motivation of RCR. Traditional Agent (Full-Context) exposes all agents to the entire shared memory, leading to redundant token usage and context flooding. RCR keeps the agent topology unchanged and allocates compact, role-relevant memory subsets under token budgets, changing what each agent sees rather than which agents are used.

Figure 2: Iterative RCR architecture with adaptive feedback loop. At each interaction round $t$ , RCR dynamically routes semantically filtered memory to each agent based on its role and current task stage. Agents process the routed contexts and produce structured outputs, which are integrated into the updated Shared Memory $M_{t+1}$ through the Memory Update step. This iterative process forms a closed-loop cycle ( $M_{t}\!\rightarrow\!$ Role-Aware Routing $\!\rightarrow\!$ Agents $\!\rightarrow\!$ IRF $\!\rightarrow\!M_{t+1}$ ), enabling progressive refinement of contextual relevance and efficient multi-agent coordination.

Benchmark	Method	Avg Runtime (s) $\downarrow$	Token (K) $\downarrow$	Primary Metric $\uparrow$
HotPotQA (F1, 0–100)	Full-Context	150.65	5.10	74.6 $\pm$ 0.6
	Static Routing	128.29	3.85	76.1 $\pm$ 0.6
	Self-RAG	118.44	3.92	78.4 $\pm$ 0.5
	End-to-End Retrieval	110.12	3.80	79.0 $\pm$ 0.5
	Multi-Agent Debate	105.33	4.06	80.0 $\pm$ 0.4
	RCR (Ours)	91.58	3.63	83.6 $\pm$ 0.5

Table 1: Main results on HotPotQA, ALFWorld, and WebShop. Each benchmark is evaluated with its official primary metric: F1 score for HotPotQA, task success rate (%) for ALFWorld, and reward score (0–1) for WebShop. HotPotQA uses three roles, while ALFWorld and WebShop use four-role configurations as detailed in Appendix F . Runtime and token usage are mean values over five independent runs; primary metrics are mean

\pm

std. Full Precision/Recall breakdowns for HotPotQA are provided in Appendix G.2 .

Method	HotPotQA			ALFWorld			WebShop
Method	Tok. $\downarrow$	Run. $\downarrow$	F1 $\uparrow$	Tok. $\downarrow$	Run. $\downarrow$	Succ. $\uparrow$	Tok. $\downarrow$	Run. $\downarrow$	Reward $\uparrow$
BM25 Retrieval	$3.79{\pm}0.05$	$103.4{\pm}1.8$	$78.2{\pm}0.5$	$4.44{\pm}0.06$	$108.2{\pm}2.1$	$63.1{\pm}0.7$	$6.12{\pm}0.07$	$123.6{\pm}2.4$	$0.617{\pm}0.007$
Semantic Retrieval	$3.76{\pm}0.04$	$99.7{\pm}1.6$	$79.2{\pm}0.5$	$4.40{\pm}0.05$	$104.7{\pm}1.9$	$64.0{\pm}0.6$	$6.05{\pm}0.06$	$119.4{\pm}2.1$	$0.624{\pm}0.006$
Role-only Routing	$3.71{\pm}0.04$	$96.8{\pm}1.5$	$80.5{\pm}0.4$	$4.36{\pm}0.05$	$100.3{\pm}1.7$	$65.9{\pm}0.6$	$5.99{\pm}0.06$	$115.6{\pm}1.9$	$0.640{\pm}0.006$
Stage-only Routing	$3.74{\pm}0.04$	$98.5{\pm}1.6$	$79.7{\pm}0.5$	$4.38{\pm}0.05$	$101.4{\pm}1.8$	$65.2{\pm}0.7$	$6.01{\pm}0.06$	$117.1{\pm}2.0$	$0.633{\pm}0.006$
Fixed Budget	$3.69{\pm}0.03$	$95.4{\pm}1.4$	$81.2{\pm}0.5$	$4.34{\pm}0.04$	$98.9{\pm}1.6$	$66.3{\pm}0.6$	$6.16{\pm}0.07$	$112.8{\pm}1.8$	$0.629{\pm}0.006$

Table 2: Controlled routing comparison under identical agents, prompts, interaction rounds, memory construction, and stopping criteria. Only the memory selection policy differs. We report mean

\pm

standard deviation over five independent runs. RCR improves the cost–quality trade-off over controlled routing variants; gains over Fixed Budget are strongest on HotPotQA and WebShop, and moderate on ALFWorld.

Benchmark	T	Avg Rt (s) $\downarrow$	Tok (K) $\downarrow$	Metric $\uparrow$
HotPotQA (F1, 0–100)	1	91.6	3.73	83.6
	2	115.8	5.92	85.0
	3	141.7	7.68	85.8
ALFWorld (Success Rate, %)	1	94.5	4.30	67.3
	2	118.5	6.28	70.9
	3	143.9	7.85	71.7

Table 3: Effect of Iterative Routing Feedback (IRF) across HotPotQA , ALFWorld , and WebShop . Each benchmark is evaluated using its official primary metric. Increasing routing iterations

T

improves task performance but increases runtime and token usage.

Variant	Rt (s) $\downarrow$	Tok (K) $\downarrow$	F1 $\uparrow$
Single-pass RCR ( $T{=}1$ )	91.6	3.73	83.6
Static memory rerouting ( $T{=}2$ )	109.4	5.48	84.1
Extra rounds w/o update ( $T{=}2$ )	112.7	5.61	83.5
Full-Context same rounds ( $T{=}2$ )	196.4	8.74	81.2
IRF w/ memory update ( $T{=}2$ )	115.8	5.92	85.0

Table 4: Controlled comparison of iterative routing feedback and additional computation on HotPotQA. All

T{=}2

variants use comparable model-call budgets.

Setting	Traj. Len	Mem Gro	Tok (K) $\downarrow$	Rt (s) $\downarrow$
Normal	1x	1x	3.63	91.6
+Retry-1	1.25x	1.25x	4.12	103.5
+Retry-2	1.5x	1.5x	4.67	116.8

Table 5: Failure-recovery analysis on HotPotQA.

Component	Latency (s)	Percentage
Scorer	1.2	1.3%
TBA + Memory Selection	0.8	0.9%
Agent LLM Inference	89.6	97.8%
Total	91.6	100%

Table 6: RCR’s routing overhead on HotPotQA. .

Figure 3: Agent-count sensitivity on HotPotQA as agent count increases ( $N\in\{3,4,5,6\}$ ). All points are measured under the same evaluation split, prompts, stopping rules, and routing implementation as the main experiments; the dashed vertical line marks the default $N{=}3$ configuration from Table 1 . (Left) Token usage increases with $N$ ; the gray dashed $\mathcal{O}(N|M|)$ line is a reference trend, and RCR grows more slowly than Full-Context. (Center) Runtime follows a similar pattern. (Right) RCR maintains stable or improved F1 as auxiliary roles are added, while Full-Context shows no consistent gain under uniform context exposure.

Figure 4: Distribution of selected memory types across roles on HotPotQA. This pattern is consistent with role-conditioned routing: the Planner receives predominantly reasoning traces and sub-plans, while the Searcher receives mostly retrieved evidence, suggesting that RCR allocates functionally distinct memory subsets rather than merely shortening prompts uniformly.

Benchmark	Configuration	Avg Runtime (s) $\downarrow$	Token (K) $\downarrow$	Primary Metric $\uparrow$
HotPotQA (F1)	Full (TBA + IS + SFR)	91.58	3.63	83.6
	w/o TBA	101.8	3.92	79.4
	w/o IS (Recency only)	96.9	3.84	78.0
	w/o SFR (FC-in-B)	108.7	3.77	76.5
ALFWorld (Success %)	Full (TBA + IS + SFR)	94.5	4.30	67.3
	w/o TBA	107.3	4.56	64.4

Table 7: Module-wise ablations under the default single-pass setting (

T{=}1

). TBA = Token Budget Allocator, IS = Importance Scorer, and SFR = Semantic Filter and Routing. Each benchmark is evaluated using its official primary metric. Removing any module degrades the efficiency–performance trade-off.

Benchmark	Strategy	Avg Runtime (s) $\downarrow$	Token (K) $\downarrow$	Primary Metric $\uparrow$
HotPotQA (F1)	MMR Retrieval	94.8 $\pm$ 1.2	3.72 $\pm$ 0.06	81.8 $\pm$ 0.6
	Learned Reranker	93.6 $\pm$ 1.1	3.70 $\pm$ 0.05	82.5 $\pm$ 0.5
	Prompt Compression	89.7 $\pm$ 1.0	3.10 $\pm$ 0.04	81.5 $\pm$ 0.7
	RCR	91.6 $\pm$ 1.3	3.63 $\pm$ 0.05	83.6 $\pm$ 0.5 ∗
ALFWorld (Success %)	MMR Retrieval	101.2 $\pm$ 1.8	4.40 $\pm$ 0.08	66.7 $\pm$ 0.8
	Learned Reranker	99.8 $\pm$ 1.6	4.37 $\pm$ 0.07	67.0 $\pm$ 0.7

Table 8: Comparison with generic context optimization strategies under identical settings. Runtime/token usage are mean

\pm

std over five runs; ∗ denotes significant improvement over the strongest baseline (

p<0.05

Level	Role Decom	Rol	Rt (s) $\downarrow$	Tok $\downarrow$	F1 $\uparrow$
Coarse	Reasoner; Executor	2	93.1	3.75	82.4
Default	Planner; Searcher; Rec	3	91.6	3.63	83.6
Fine	Planner; Searcher; Checker; Composer; Verifier	5	96.2	4.02	83.8

Table 9: Role granularity sensitivity on HotPotQA.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Train/Val/Test	Router Traj.	Test Inst.	Mem Items	Mem Tokens
HotPotQA	90,447 / 7,405 / 7,405 (fullwiki)	1,000	7,405	18.6	3.2K
ALFWorld	3,553 / 140 / 134	800	134	24.3	4.1K
WebShop	10,587 / 1,000 / 500	1,000	500	31.7	5.6K

Appendix

Table 10: Dataset statistics and routing data configuration.

$\lambda_{\mathrm{Tok}}$	$\lambda_{\mathrm{Lat}}$	F1 $\uparrow$	Token (K) $\downarrow$	Runtime (s) $\downarrow$
0.01	0.01	83.4	3.91	96.8
0.05	0.01	83.6	3.63	91.6
0.10	0.01	83.1	3.41	88.9
0.05	0.05	83.2	3.58	89.7
0.05	0.10	82.7	3.49	87.5

Appendix

Table 11: Sensitivity to reward coefficients on the HotPotQA development split. The selected setting is marked in bold. Token and runtime are measured from actual execution.

Component	Specification
Agent interface	Same role descriptions, task prompts, tool interfaces, round limits, and stopping rules are used across controlled multi-agent methods.
Role metadata	Roles are treated as metadata provided by the underlying agent framework or decomposition module. RCR does not require a fixed role taxonomy.
Stage metadata	Stage descriptors are used as optional routing metadata. Baselines that do not use stage-conditioned routing ignore this metadata.
Memory pool	All controlled methods use the same shared-memory pool within each benchmark.
Routing policy	The only varied component is how memory items are selected and routed to agents.
Evaluation	QA benchmarks use Precision/Recall/F1. ALFWorld and WebShop use their benchmark-specific success or task-score metrics. Runtime and token usage are reported across all benchmarks.

Appendix

Table 12: Reproducibility checklist for controlled experiments.

Benchmark	Traj.	Items	Pos.	Hard Neg.	Soft Neg.	Unlabeled
HotPotQA	1,000	48.2K	18.6%	31.4%	17.3%	32.7%
ALFWorld	800	39.5K	16.9%	29.8%	16.6%	36.7%
WebShop	1,000	56.7K	15.4%	33.1%	17.0%	34.5%

Appendix

Table 13: Weak-label statistics for training the importance scorer. We use a three-way triage: high-confidence positives, high-confidence negatives, soft negatives (reduced loss weight

\eta

), and unlabeled ambiguous items (zero loss weight).

Benchmark	Memory item	Label type	Reason
HotPotQA	Retrieved passage cited by the Searcher and later used by the Synthesizer in a successful answer trajectory.	Positive	Linked to the logged evidence chain used in the successful trajectory.
HotPotQA	Entity-overlapping passage retrieved for the same question but contradicted by the later accepted evidence.	Hard negative	Same-role–stage candidate whose content conflicts with the logged successful evidence.
HotPotQA	Unused passage with partial entity overlap but no clear support or contradiction in the trajectory.	Unlabeled	Relevance cannot be reliably inferred; masked out from the scorer loss to avoid false-negative noise.
ALFWorld	Action/state trace: open fridge; put apple in fridge .	Positive	Appears in a successful logged action trajectory satisfying the goal condition.
ALFWorld	Action/state trace: put apple on counter when the target goal requires storing the apple in the fridge.	Hard negative	Contradicts the target state observed in the successful trajectory.
ALFWorld	Unused navigation observation from a different room with weak lexical overlap to the goal.	Soft negative	Likely irrelevant due to role–stage mismatch, but not contradicted with full confidence; included with reduced loss weight.

Appendix

Table 14: Examples of weak-label construction for the importance scorer. Labels are derived from logged training trajectories without gold evidence annotations, test trajectories, or future-round leakage.

Benchmark	Backbone	Method	Avg Runtime (s) $\downarrow$	Token (K) $\downarrow$	Primary Metric $\uparrow$
HotPotQA (F1, 0–100)	GPT-5-mini	Full-Context	150.65	5.10	74.6 $\pm$ 0.6
	GPT-5-mini	Static Routing	128.29	3.85	76.1 $\pm$ 0.6
	GPT-5-mini	RCR	91.58	3.63	83.6 $\pm$ 0.5
	Qwen3-Max	Full-Context	162.8	5.07	70.9 $\pm$ 0.8
	Qwen3-Max	Static Routing	137.6	3.86	72.4 $\pm$ 0.7
	Qwen3-Max	RCR	101.7	3.64	78.6 $\pm$ 0.7

Appendix

Table 15: Backbone-transfer robustness across model families and deployment settings. The Importance Scorer and Token Budget Allocator are trained once using GPT-5-mini trajectories and then frozen when evaluated with Qwen3-Max and Gemini 2.5 Pro agents. Qwen3-Max provides an open-weight 7B backbone, while Gemini 2.5 Pro provides a stronger closed-source API backbone. This experiment tests whether RCR transfers across model families, capability levels, and open/closed deployment settings without backbone-specific retraining. Results are intended as a robustness check rather than a full re-evaluation of all baselines.

Benchmark	Method	Avg Runtime (s) $\downarrow$	Token (K) $\downarrow$	Precision $\uparrow$	Recall $\uparrow$	F1 $\uparrow$
HotPotQA	Full-Context	150.65	5.10	72.3	75.1	74.6
	Static Routing	128.29	3.85	74.8	77.5	76.1
	Self-RAG	118.44	3.92	77.5	79.3	78.4
	End-to-End Retrieval	110.12	3.80	78.0	80.1	79.0
	Multi-Agent Debate	105.33	4.06	79.1	81.0	80.0
	RCR (Ours)	91.58	3.63	81.8	84.4	83.6

Appendix

Table 16: Additional results on HotPotQA , MuSiQue , and 2WikiMultihop . Token usage reflects dataset-specific memory pool sizes: MuSiQue instances require longer compositional evidence chains (avg. 4-hop), resulting in larger shared memory; 2WikiMultihop involves shorter bridge passages (avg. 2-hop) with a smaller memory footprint. All results are averaged over five independent runs.

Benchmark	Budget Strategy	Runtime $\downarrow$	Token (K) $\downarrow$	Metric $\uparrow$
HotPotQA (F1, 0–100)	Uniform static $B_{i}{=}2048$	95.4	3.69	81.2
	Dev-oracle static per-role	99.8	4.21	83.8
	Learned TBA (RCR)	91.6	3.63	83.6
ALFWorld (Success Rate, %)	Uniform static $B_{i}{=}2048$	98.9	4.34	66.3
	Dev-oracle static per-role	103.6	4.86	67.6
	Learned TBA (RCR)	94.5	4.30	67.3

Appendix

Table 17: Diagnostic comparison of learned TBA against a development-set oracle-tuned static per-role budget. Uniform static

B_{i}{=}2048

is included as a reference. The oracle baseline selects one fixed budget per role on the development split and applies it to all test instances.

Benchmark	$B_{i}$	Ave Runtime (s) $\downarrow$	Token (K) $\downarrow$	Primary Metric $\uparrow$
HotPotQA (F1)	512	75.1	2.42	71.1
	1024	86.7	3.12	76.8
	2048	91.6	3.63	83.6
	4096	126.4	5.85	82.0
ALFWorld (Success %)	512	71.3	2.21	60.5
	1024	83.8	3.13	64.6

Appendix

Table 18: Fixed-budget sensitivity analysis for per-agent budget

B_{i}

. RCR is evaluated with

B_{i}\in\{512,1024,2048,4096\}

while keeping the routing policy fixed. Each benchmark is evaluated using its official primary metric. Performance improves up to

B_{i}{=}2048

and then saturates, supporting

B_{i}{=}2048

as the reference budget level.

	Token (K) $\downarrow$				Runtime (s) $\downarrow$				F1 $\uparrow$
Method	$N{=}3$	$N{=}4$	$N{=}5$	$N{=}6$	$N{=}3$	$N{=}4$	$N{=}5$	$N{=}6$	$N{=}3$	$N{=}4$	$N{=}5$	$N{=}6$
Full-Context	5.10	6.82	8.49	10.18	150.7	201.2	251.3	301.8	74.6	75.8	75.2	74.3
Static Routing	3.85	5.14	6.40	7.65	128.3	171.4	213.8	256.2	76.1	77.2	76.9	76.5
RCR (Ours)	3.63	4.53	5.31	5.98	91.6	109.5	124.3	137.6	83.6	84.8	85.3	85.7

Appendix

Table 19: Scalability on HotPotQA as agent count grows from

N{=}3

N{=}6

by adding Verifier, Critic, and Refiner roles.

N{=}3

(Planner / Searcher / Recommender) is identical to the main experiment (Table 1 ). Token in thousands; runtime in seconds.

Role	Stage	Time	Memory item
Planner	Decompose	$t{=}1$	Sub-query 1: find the oldest university in the English-speaking world.
Planner	Decompose	$t{=}1$	Sub-query 2: find rivers sharing a name with a U.S. state.
Searcher	Evidence	$t{=}2$	Oxford University was founded around 1096–1167 and is located in Oxford, England.
Searcher	Evidence	$t{=}2$	River Thames flows through Oxford; River Mississippi shares a name with a U.S. state.
Searcher	Evidence	$t{=}2$	River Ohio flows through Oxford, Ohio, giving a candidate river–state match.
Searcher	Stale	$t{=}1$	[Stale] Cambridge was also considered; this is contradicted by later evidence.

Appendix

Table 20: Shared memory pool for the qualitative routing example. All items are visible to Full-Context agents, while RCR selects role-specific subsets under each agent’s token budget.

Planner receives	Searcher receives	Recommender receives
Sub-query 1: oldest English-speaking university.	Oxford University was founded around 1096–1167 and is located in Oxford, England.	Oxford University was founded around 1096–1167 and is located in Oxford, England.
Sub-query 2: river sharing a U.S. state name.	River Thames flows through Oxford; River Mississippi shares a name with a U.S. state.	River Ohio flows through Oxford, Ohio, giving a candidate match.
Reasoning trace: verify whether the candidate city satisfies both constraints.	River Ohio flows through Oxford, Ohio, giving a candidate match.	Synthesis note: the Oxford candidate needs revision on the river-name condition.
Filtered item: the stale Cambridge item is suppressed. The Planner receives reasoning traces without raw evidence overload; the Searcher receives evidence-heavy memory; and the Recommender receives evidence plus synthesis signals rather than all planning traces.
Resulting outputs: The Planner decomposes the task into university identification and river–state verification. The Searcher flags that Oxford, England satisfies the university condition, while Oxford, Ohio provides the river–state match. The Recommender revises the initial candidate and avoids relying on the stale Cambridge memory.

Appendix

Table 21: Case A: successful role-specific routing. RCR selects planning traces for the Planner, evidence-heavy items for the Searcher, and compact evidence plus synthesis notes for the Recommender.

Planner receives	Searcher receives	Recommender receives
Sub-query 1: oldest English-speaking university.	Oxford University was founded around 1096–1167 and is located in Oxford, England.	Oxford University was founded around 1096–1167 and is located in Oxford, England.
Sub-query 2: river sharing a U.S. state name.	[Stale] Cambridge was also considered.	Synthesis note: the river condition needs revision.
Reasoning trace: verify both constraints.	River Thames flows through Oxford.	Missing key evidence: River Ohio match is absent.
Dropped item: the River Ohio evidence is not routed to the Searcher because the stale Cambridge item consumes part of the budget. The Recommender therefore lacks the key river–state evidence needed to complete verification.
Resulting outputs: The Planner still produces the correct decomposition, but the Searcher reports Oxford/Cambridge as competing university candidates and fails to retrieve the River Ohio evidence. The Recommender returns an incomplete or incorrect answer because the river condition remains unresolved.

Appendix

Table 22: Case B: imperfect routing under stale-memory interference. The scorer assigns high relevance to a stale item, causing key river evidence to be dropped under the Searcher’s budget.

Memory Type	Planner	Searcher	Recommender
Reasoning traces	61%	12%	24%
Retrieved evidence	8%	71%	45%
Sub-plans	28%	9%	18%
Tool outputs	3%	8%	13%

Appendix

Table 23: Distribution of selected memory types across roles on HotPotQA. Values denote the percentage of routed memory items assigned to each role.

Agent	Mixed context window
Planner	Current goal: decompose the question into two sub-queries ; earlier brainstorming suggested Cambridge but was later contradicted ; identify the oldest English-speaking university ; duplicate raw evidence from a previous search ; verify whether the candidate city also satisfies the river-name condition ; unrelated notes about university rankings and old search attempts.
Searcher	Oxford University was founded around 1096–1167 and is located in Oxford, England ; stale snippet about Cambridge as a possible oldest university ; River Ohio flows through Oxford, Ohio and shares its name with a U.S. state ; duplicate search results about the River Thames ; retrieve evidence linking each candidate city to the relevant constraint ; broken links, unrelated pages, and outdated retrieval keywords.
Recommender	The answer must satisfy both the university condition and the river-name condition ; unused draft answer: Oxford, England ; verified evidence supports revising the initial candidate ; repeated evidence fragments and obsolete summaries ; final synthesis should use only verified evidence ; irrelevant meeting notes and superseded reasoning traces.

Appendix

Table 24: Illustrative HotPotQA Full-Context examples showing context flooding in Figure 1 . Black spans denote function-relevant content, while red spans denote redundant, stale, or role-irrelevant content interleaved within the same context window.

Agent	RCR routed context
Planner	Current goal: decompose the question into two sub-queries. Identify the oldest English-speaking university. Verify whether the candidate city also satisfies the river-name condition.
Searcher	Oxford University was founded around 1096–1167 and is located in Oxford, England. River Ohio flows through Oxford, Ohio and shares its name with a U.S. state. Retrieve evidence linking each candidate city to the relevant constraint.
Recommender	The answer must satisfy both the university condition and the river-name condition. Verified evidence supports revising the initial candidate. Final synthesis should use only verified evidence.

Appendix

Table 25: RCR routed-context example on the same HotPotQA instance. RCR suppresses stale, duplicate, and role-irrelevant items, and routes compact role-relevant subsets to each agent under token budgets.

Agent	Mixed context window
Planner	Current goal: find a lightweight running shoe under $100 ; earlier plan suggested formal leather shoes but was abandoned ; decompose the task into price, category, size, and review constraints ; duplicate notes from previous searches for hiking boots ; prioritize products matching the user’s stated budget and use case ; unrelated browsing history about office chairs and phone cases.
Searcher	Product result: lightweight running shoe, size 10, price $79.99, average rating 4.5 ; stale result: out-of-stock running shoe from an earlier page ; retrieved attribute: breathable mesh upper and suitable for daily running ; duplicate search result from the same seller with a different thumbnail ; compare price, availability, rating, and shipping constraints ; unrelated product snippets for dress shoes, sandals, and backpacks.
Recommender	The selected item should satisfy category, price, availability, and user preference constraints ; unused draft recommendation for a leather loafer ; verified evidence supports the lightweight running shoe as the best match ; repeated product descriptions and obsolete cart candidates ; final recommendation should cite price, rating, and matching attributes ; irrelevant notes about previously rejected products and unrelated categories.

Appendix

Table 26: Illustrative WebShop Full-Context examples showing context flooding in Figure 1 . Black spans denote function-relevant content, while red spans denote redundant, stale, or role-irrelevant content interleaved within the same context window.

Agent	RCR routed context
Planner	Current goal: find a lightweight running shoe under $100. Extract constraints: category = running shoe, price <$ 100, size = 10, preference = lightweight and well-reviewed. Prioritize candidates matching the user’s stated budget and use case.
Searcher	Candidate product: lightweight running shoe, size 10, price $79.99, average rating 4.5. Retrieved attributes: breathable mesh upper, daily running use, available stock. Compare price, availability, rating, and shipping constraints.
Recommender	Verified evidence supports the lightweight running shoe as the best match. The item satisfies category, price, size, availability, and user preference constraints. Final recommendation should cite price, rating, and matching attributes.

Appendix

Table 27: RCR routed-context example on the same WebShop instance. RCR removes stale, duplicate, and role-irrelevant items, and routes compact role-relevant subsets to each agent under token budgets.

Agent	Mixed context window
Planner	Current goal: put a clean mug on the coffee table ; earlier plan searched the bathroom sink but was abandoned ; decompose the task into locating a mug, cleaning it, and moving it to the target receptacle ; duplicate notes about opening already-checked drawers ; track unresolved constraints: mug must be clean and placed on the coffee table ; unrelated observations about a knife, apple, and bedroom lamp.
Searcher	Observation: a dirty mug is on the kitchen counter ; stale observation: mug might be in the cabinet, contradicted by later observation ; useful receptacles: sink can clean the mug, coffee table is the target location ; repeated action log: open drawer, close drawer, open drawer ; retrieve next actionable locations for washing and placing the mug ; unrelated room-state traces from the bedroom and pantry.
Recommender	The next action should move toward satisfying both conditions: clean the mug and place it on the coffee table ; unused draft action: put mug in cabinet ; verified trace supports the sequence: pick up mug, go to sink, clean mug, go to coffee table ; obsolete action candidates from failed drawer exploration ; final action proposal should avoid revisiting checked locations ; irrelevant object notes about the knife, apple, and lamp.

Appendix

Table 28: Illustrative ALFWorld Full-Context examples showing context flooding in Figure 1 . Black spans denote function-relevant content, while red spans denote redundant, stale, or role-irrelevant content interleaved within the same context window.

Agent	Routed context type	RCR routed context
Planner	Reasoning state and sub-plan	Current goal: put a clean mug on the coffee table. Decompose the task into: locate mug, clean mug, and move it to the coffee table. Track constraints: the mug must be clean and the final receptacle must be the coffee table.
Searcher	Evidence and actionable observations	Observation: a dirty mug is on the kitchen counter. Useful receptacle: the sink can be used to clean the mug. Target location: the coffee table. Next actionable locations: kitchen counter, sink, and coffee table.
Recommender	Verified action trace and next action	Verified trace: pick up mug, go to sink, clean mug, go to coffee table. Avoid revisiting checked drawers or unrelated rooms. Recommended next step: continue the action sequence that satisfies both the clean-mug and target-placement constraints.

Appendix

Table 29: RCR routed-context example on the same ALFWorld instance. RCR removes stale, duplicate, and role-irrelevant items, and routes compact role-relevant subsets to each agent under token budgets.

LLM Context Management

Large language model context management studies how to select, compress, retrieve, and organize information within bounded model inputs.

See all 246 papers

Aug 31, 2026

Learning What to Retain: Gated-Memory Routing for Efficient Collaboration in Multi-Agent LLM Systems

Large language model (LLM)-based multi-agent systems tackle complex reasoning by orchestrating how multiple agents are configured and how they collaborate. A central challenge is to adapt orchestration to the evolving collaboration state. Routing from the query alone cannot adapt to intermediate progress or errors, which hurts accuracy. Routing from the complete execution history supplies this missing context, but forces later decisions to process every prior step, including redundant or low-utility ones. This creates an execution-history overload that inflates cost. Effective orchestration instead requires a compact state that captures useful progress without accumulating redundant context. We propose Gated-Memory Routing, which conditions each decision on the query and a learned execution memory. A learned Memory Write Gate commits only non-redundant reasoning steps, and a learned Retrieval Gate supplies each agent a compact, relevant subset, so every decision conditions on a clean, informative state. At each step, the system selects the next role and backbone from this memory, while an Adaptive Halting Controller stops execution once the memory contains sufficient evidence for answering. Across five reasoning and code-generation benchmarks, our framework is both effective and efficient: it attains the best average accuracy, exceeding the strongest baseline by 2.44 points, while reducing HumanEval inference cost by 31.9% relative to that baseline. Code is available at https://github.com/rajibrhasan/gated-memory-routing

Rakibul Hasan Rajib, Mengxin Zheng, Qian Lou

University of Central Florida

Multi-Agent LLM Systems Multi-Agent System Optimization LLM Agent Orchestration LLM Agent Workflow Optimization LLM Agents LLM Routing+1

Ask PDF

Apr 24, 2026

ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents

Large language model (LLM) agents often struggle in long-context interactions. As the history grows, sliding windows and prompt compression may drop earlier information that later steps rely on. Retrieval-based memory systems surface relevant content but still overlook the causal and logical dependencies between steps. We introduce ContextWeaver, a selective and dependency-structured memory framework that organizes an agent's interaction trace into a graph of reasoning steps and selects the relevant context for future actions. ContextWeaver supports (1) dependency-based construction and traversal that link each step to the earlier steps it relies on; (2) compact dependency summaries of the reasoning paths; and (3) a lightweight validation layer that incorporates execution feedback. On five software-engineering and terminal benchmarks with a fixed budget per task, ContextWeaver improves the resolve rate over a sliding-window baseline on all five, with a significant gain when pooled across benchmarks, while using 13-18% fewer agent steps. It also performs favorably against alternative context-management approaches, with the largest gains on long-horizon repository tasks. On SWE-Bench Pro, ContextWeaver raises the resolve rate from 66.5% to 72.6%, with significant improvements across all three model backbones. The improvement also holds when the sliding window's budget is raised to cover the summary cost.

Yating Wu, Yuhao Zhang, Sayan Ghosh +4

UT Austin · AWS AI Labs

LLM Agent Memory Graph-Based Agent Memory LLM Context Management Agent Memory Management

Ask PDF

May 28, 2026

Enhancing Multi-Agent Communication through Attention Steering with Context Relevance

LLM-based multi-agent systems have demonstrated remarkable performance on complex tasks through collaborative reasoning. However, these systems tend to rapidly accumulate extremely long conversation histories during interaction. As conversations lengthen, relevant information is increasingly diluted by irrelevant context, leading to degraded performance. In this work, we present Agent-Radar, a training-free context management method that dynamically steers each agent's attention toward relevant context with a novel temporal and spatial decay mechanism. Our experiments demonstrate that Agent-Radar outperforms state-of-the-art methods across five different benchmarks, yielding gains of up to 7.64 absolute points. Furthermore, our analysis shows that Agent-Radar remains effective and robust as the number of agents and interaction rounds increases. Finally, the ablation study shows that core components in Agent-Radar are crucial to performance and generalizable in different settings.

Hongxiang Zhang, Yuan Tian, Tianyi Zhang

Purdue University West Lafayette, IN, USA

Multi-Agent LLM Systems Multi-Agent System Optimization Multi-Agent Communication LLM Context Management

Ask PDF

Abstract

LLM Context Management

Explore similar work

Learning What to Retain: Gated-Memory Routing for Efficient Collaboration in Multi-Agent LLM Systems

ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents

Enhancing Multi-Agent Communication through Attention Steering with Context Relevance