Adaptive Memory Index Evolution for Agents

Yuntong Hu; Tongli Su; Yuanfu Sun; Shuo Ma; Qiaoyu Tan; Liang Zhao

cs.LGOct 26, 2026

Adaptive Memory Index Evolution for Agents

Authors: Yuntong Hu, Tongli Su, Yuanfu Sun, Shuo Ma, Qiaoyu Tan, Liang Zhao

Organizations: Emory University · New York University · Meituan

Abstract

Long-horizon agents continually accumulate experience, yet retaining a memory does not guarantee access when it is needed. Existing systems can adapt both memory content and access; we isolate persistent access updates as a separately controlled adaptation layer. We formulate this target as Memory Index Evolution and instantiate it with Adaptive Memory Index Evolution (AMIE), a lightweight plug-in layer that evolves per-memory access state from deployment retrievals without modifying stored content or retraining the host memory system. AMIE distils a semantic teacher's keep/suppress judgments into a retrieval-conditioned policy and accumulates its selective actions into bounded state that reweights subsequent host retrievals. Surprisingly, even a zero-call policy that simply accumulates retrieval frequency can match or outperform semantic policies in several settings, revealing retrieval traffic itself as a strong signal for access adaptation, while learned selectivity improves robustness when frequently retrieved memories remain useful. Across five retrieval-based hosts and three benchmarks, trained AMIE improves downstream F1 or accuracy in all 15 comparisons, by 0.70--3.37 percentage points (up to 25% relative), and improves retrieval in all 10 comparisons. Controlled analyses show that persistent competitor suppression can expose previously unseen evidence, while suppressing repeatedly useful memories creates the complementary failure mode. Together, these results establish persistent memory access as a practical adaptation layer, distinct from the content it serves, in retrieval-based agent memory systems.

Figures & tables

Figure 1: Persistent index adaptation improves both memory access and downstream utility across heterogeneous memory systems. We evaluate three complementary memory settings: LoCoMo for long-horizon conversation, GroupMem for cross-channel group memory, and MemBench for personal memory. Trained AMIE improves all downstream host–benchmark comparisons and all comparisons with retrieval measurements. (a,b) Absolute LoCoMo F1 and GroupMem accuracy, with relative gains over each matched host shown at the vertices. (c) Relative retrieval gains on LoCoMo and MemBench. (d) All ten pairs measured at both levels improve in both retrieval and downstream utility, showing that the benefits of index evolution extend beyond retrieval metrics to the tasks memory ultimately serves.

Panel of Figure 2 — Figure 2: AMIE applies persistent access state to subsequent retrievals. Native candidates supply both the answer context, ranked with pre-update state, and the policy’s feedback. Dashed arrows show state access and the commit constraint: updates occur after context sealing and affect later queries.

			+AMIE
Benchmark	Host	Host only	Trained	Frozen	Count
LoCoMo ( Maharana et al., 2024 )	BM25 ( Robertson and Zaragoza, 2009 )	51.94	$53.02_{+1.08}$	$53.21_{+1.27}$	$53.41_{+1.47}$
	Mem0 ( Chhikara et al., 2025 )	42.17	$44.42_{+2.25}$	$44.28_{+2.11}$	$45.03_{+2.86}$
	MemoryOS ( Kang et al., 2025 )	45.29	$45.47_{+0.18}$	$46.20_{+0.91}$	$45.90_{+0.61}$
	A-MEM ( Xu et al., 2025 )	63.27	$65.14_{+1.87}$	$64.73_{+1.46}$	$66.02_{+2.75}$
	Memory-R2 ( Yan et al., 2026a )	65.92	$67.78_{+1.86}$	$66.57_{+0.65}$	$65.03_{-0.89}$

Table 1: Retrieval scores with and without AMIE (%, higher is better). Subscripts subtract the displayed baseline; green/red intensity encodes positive/negative magnitude, not significance. Host-specific retrieval units are defined in Appendix A.2 .

			+AMIE
Benchmark	Host	Host only	Trained	Frozen	Count
LoCoMo (F1) ( Maharana et al., 2024 )	No memory	0.00	-	-	-
	BM25 ( Robertson and Zaragoza, 2009 )	17.73	$19.04_{+1.31}$	$18.40_{+0.67}$	$18.27_{+0.54}$
	Mem0 ( Chhikara et al., 2025 )	9.98	$12.44_{+2.46}$	$12.69_{+2.71}$	$12.85_{+2.87}$
	MemoryOS ( Kang et al., 2025 )	20.27	$20.97_{+0.70}$	$20.83_{+0.56}$	$20.77_{+0.50}$
	A-MEM ( Xu et al., 2025 )	16.54	$17.43_{+0.89}$	$17.17_{+0.63}$	$16.92_{+0.38}$

Table 2: Downstream utility with and without AMIE (%, higher is better). F1 is used for LoCoMo and accuracy for GroupMem and MemBench. Subscripts and green/red shading encode signed percentage-point differences from the displayed baselines, not significance; dashes mark inapplicable arms.

Figure 3: Competition and state interventions in the BM25 replay. Panels compare (a) added-store content, (b) competition and reachable headroom, and (c) oracle and direction controls. Recall-point differences cover $1{,}527$ questions, with conversation-cluster $95\%$ intervals where records exist; baselines and controls are defined in Appendix C and Appendix F.1 .

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Benchmark scope (util./retr.)	Memory units	Setting	Metric	Revisiting
LoCoMo	1,540 / 1,527	5,882 turns (10 conv.)	conversational	F1	yes
GroupMem	745 / –	120,000 messages (33 channels)	cross-channel workplace	Acc. (GPT-5)	yes
MemBench	23,137 / 23,132	${\sim}1.2$ M memories	personal	Acc. (MCQ)	history retrieval

Appendix

Table A.1: Benchmark statistics and evaluation scope. Counts describe the benchmark-level utility and evidence-valid scopes, not completed sample sizes for every host; retrieval requires resolvable supporting-evidence identifiers. Revisiting describes repeated access to a store, not a measured guarantee of competitor recurrence.

Panel of Figure 7 — Figure A.1: Four native-index attachment examples. Grey modules retain host behaviour, blue modules reweight candidates, and red feedback updates per-memory state for later queries. The integration settings are specified in Table A.2 .

Host	Native index / unit	Score adapter	Attachment / depth	Following stage
BM25	Inverted index / turn	Nonnegative BM25	Scored turn list / $10$	Turn delivery
Mem0	Fact vectors and user filter / fact	Squared distance $d\mapsto 1/(1+d)$	Filtered fact list / $10$	Fact delivery
MemoryOS	Topic and page vectors / page	Inner product; segment and page gates $0.1$	Scored pages / $7$	Merge history, profile, knowledge
A-MEM	Enriched-note vectors and links / root	Cosine $s\mapsto(1+s)/2$	Root list / $10$	Link expansion; $1{,}200$ -word cap

Appendix

Table A.2: Host interfaces for persistent access updates. AMIE reweights the scored candidates at the listed attachment point; subsequent native expansion can propagate a root’s suppression to linked evidence.

Policy	Cost / query	Representative effect	Role in the model
Trained	$87$ ms	MemoryOS F1 $+0.70$ ; Mem0 F1 $+2.46$	Learned selective actions
Frozen	$24$ s	MemoryOS recall $+0.91$	Direct semantic mask inference
Count	$0$ model calls	Mem0 F1 $+2.87$ ; BM25 R@10 $+1.47$	Constant all-ones action mask

Appendix

Table A.3: Policies sharing the AMIE action and state interface. Representative effects follow Table 1 and Table 2 ; serving costs follow Table G.1 .

Host	Policy	w/o AMIE	+AMIE	$\Delta$	95% CI
Mem0	trained	9.98	12.44	+2.46	$[+1.60,+3.24]$
	frozen	9.98	12.69	+2.71	$[+1.73,+3.53]$
	count	9.98	12.85	+2.87	$[+1.95,+3.80]$
MemoryOS	trained	20.27	20.97	+0.70	$[+0.29,+1.14]$
	frozen	20.27	20.83	+0.56	$[+0.10,+1.01]$
	count	20.27	20.77	+0.50	$[-0.13,+1.05]$

Appendix

Table B.1: Reproducible LoCoMo utility pairs. Scores and displayed

\Delta

follow Table 2 ; intervals use the unrounded per-question differences and conversation-cluster

95\%

bootstrap.

Table 11

Figure C.1: Controlled state interventions on BM25 and MemoryOS. (a) BM25 oracle and direction controls; (b) windows and decay; (c) resets and order comparisons. Recall-point differences use conversation-cluster $95\%$ intervals for per-question comparisons and paired $t$ intervals for fifty-order comparisons; missing intervals are omitted. Appendix C defines the controls and their baselines.

Figure C.2: Policy effects depend on evidence exposure in the LoCoMo replay. Panels (a,b) stratify BM25 and MemoryOS by prior gold exposure; (c) stratifies count effects by candidate exposure. Whiskers show conversation-cluster

95\%

intervals for recall-point differences.

Figure D.1: Persistent state interventions in the BM25 replay. Panels show (a) label-informed count injection on $1{,}527$ questions, (b) frozen-state evaluation on $386$ future questions, and (c) matched-minus-swapped state on $381$ category-split and $387$ interleaved future questions. Bands and whiskers are conversation-cluster $95\%$ intervals on recall-point differences; injection is a diagnostic intervention, not a runtime policy.

$n$	competitors (non-gold top ten)	gold entries of the top ten	ranks 11–20	random pool
$1$	$+2.10$ $[+1.57,+2.68]$	$-6.17$ $[-7.12,-5.43]$	$0.00$	$0.00$
$2$	$+2.93$ $[+2.20,+3.80]$	$-10.49$ $[-11.94,-9.28]$	$0.00$	$0.00$
$4$	$+4.29$ $[+3.40,+5.19]$	$-15.40$ $[-16.53,-14.19]$	$0.00$	$0.00$
$8$	$+4.97$ $[+4.13,+5.87]$	$-21.69$ $[-22.94,-20.26]$	$0.00$	$0.00$
$16$	$+5.14$ $[+4.28,+6.05]$	$-26.51$ $[-28.13,-24.69]$	$0.00$	$0.00$
$32$	$+5.35$ $[+4.45,+6.28]$	$-29.86$ $[-31.45,-28.16]$	$0.00$	$0.00$

Appendix

Table D.1: Causal traffic injection (

\Delta

Recall@10 against the same question with an empty state, conversation-cluster

95\%

intervals,

n{=}1{,}527

). Only the count of the target set is changed; query, host scores, gold and store are identical across columns. Under the online rule the competitors’ natural median count is

2

and the gold entries’ is

1

$f$	questions observed (total)	Recall@10	$\Delta$ [95% CI]
$0$	$0$	$65.67$	$0.00$
$0.01$	$10$	$65.67$	$0.00$
$0.05$	$59$	$66.97$	$+1.30$ $[+0.28,+2.38]$
$0.10$	$114$	$67.62$	$+1.94$ $[+0.53,+3.22]$
$0.25$	$288$	$68.39$	$+2.72$ $[+1.34,+3.95]$
$0.50$	$576$	$68.01$	$+2.33$ $[+0.12,+4.20]$

Appendix

Table D.2: Traffic sufficiency on the fixed future set (

386

questions;

\Delta

Recall@10 against the unmodified host, conversation-cluster

95\%

intervals). The state is frozen after observing the first fraction

f

of each conversation’s accumulation segment; the online row keeps updating inside the future set.

$E_{k}$	gold exposure	$n$	count	frozen	trained	count $-$ frozen [95% CI]
$\geq 4$	$0$	$439$	$+5.33$	$+3.11$	$+3.65$	$+2.22$ $[+0.79,+3.90]$
$\geq 4$	$1$ – $2$	$559$	$+2.00$	$+2.90$	$+1.71$	$-0.89$ $[-1.82,+0.00]$
$\geq 4$	$3$ – $9$	$345$	$-2.89$	$-2.73$	$-1.96$	$-0.16$ $[-1.67,+1.05]$
$\geq 4$	$\geq 10$	$46$	$-4.16$	$-2.96$	$-4.96$	$-1.20$ $[-4.35,+0.27]$
$\leq 3$	any	$138$	$-0.07$	$+0.24$	$+0.05$	intervals include zero

Appendix

Table D.3: Count against the semantic arms across competitor and gold exposure (BM25 replay,

\Delta

Recall@10 against the host, conversation-cluster

95\%

intervals on the paired count-minus-frozen difference).

E_{k}

is the number of the host’s ten candidates earlier traffic had already retrieved; gold exposure is the number of earlier appearances of the question’s own gold turns.

Repeat occupancy	no gold missing	gold reachable (11–100)	gold beyond the pool
low ( $n{=}138$ )	$0.00$ ( $n{=}34$ )	$+0.61$ $[+0.00,+1.82]$ ( $n{=}55$ )	$-1.02$ $[-3.26,+0.00]$ ( $n{=}49$ )
mid ( $n{=}454$ )	$-1.93$ $[-3.78,+0.00]$ ( $n{=}207$ )	$+3.36$ $[+0.79,+6.09]$ ( $n{=}135$ )	$-0.74$ $[-1.53,+0.00]$ ( $n{=}112$ )
high ( $n{=}935$ )	$-2.78$ $[-3.59,-1.86]$ ( $n{=}485$ )	$+15.99$ $[+12.73,+19.00]$ ( $n{=}230$ )	$-0.11$ $[-0.38,+0.00]$ ( $n{=}220$ )

Appendix

Table D.4: Index evolvability map (online count rule, BM25 replay,

\Delta

Recall@10 against the host, conversation-cluster

95\%

intervals). Rows are repeat-occupancy bins of the host’s top ten (

<0.4

0.4

–

0.8

\geq 0.8

); columns class the question’s missing gold as already in the top ten, inside the pool at ranks

11

–

100

, or beyond the pool.

Arm	$\Delta$ [95% CI]	online gain / reference gain
online count rule	$+1.48$ $[+0.70,+2.21]$	—
hindsight $\rho\eta$ oracle (saved)	$+2.24$	$66\%$
hindsight $\eta$ -only oracle (saved)	$+3.13$	$47\%$
oracle top-ten suppression	$+5.54$ $[+4.61,+6.46]$	$27\%$
ideal pool recall	$+20.43$ $[+18.44,+22.61]$	$7\%$
perfect-recall reference	$+48.06$ $[+46.34,+49.81]$	$3\%$

Appendix

Table D.5: Oracle references for index-side recall gains. The top-ten intervention suppresses only original non-gold top-ten entries; pool and perfect-recall references have broader oracle access. Values are BM25 replay differences with conversation-cluster

95\%

intervals, not an optimised functional-form bound.

Figure D.2: Selectivity, oracle references, and decay. Panels (a,b) show BM25 replay recall differences with conversation-cluster

95\%

intervals; (c) compares decay to unbounded counting across fifty orders with

95\%

t

intervals. Oracle top-ten suppression is a restricted intervention, not a state-space optimum.

Split	Arm	$n$	$\Delta$ [95% CI]
categories (different deployments)	category 1, own state	$71$	$+1.76$ $[+0.01,+3.65]$
	category 2, own state	$77$	$+5.20$ $[+0.00,+11.77]$
	category 4, own state	$213$	$+3.05$ $[+0.50,+5.48]$
	paired: own $-$ other deployment’s state	$381$	$\mathbf{+2.16}$ $[+0.54,+3.90]$
	paired: own $-$ union of the others	$381$	$+1.12$ $[-1.16,+4.10]$
even / odd (same distribution)	even, own state	$194$	$+2.58$ $[+0.00,+5.20]$

Appendix

Table D.6: Content–index swap (BM25 replay, frozen states,

\Delta

Recall@10 against the unmodified host on each deployment’s future set, conversation-cluster

95\%

intervals). The paired rows pool every future question of a split and compare its own state with a state evolved on another deployment of the same store.

Unit	units	off	+count	purity	$\Delta$ [95% CI]	improved	worsened
1 turn	$5{,}882$	$51.94$	$53.41$	$1.00$	$+1.48$ $[+0.70,+2.21]$	$57$	$30$
2 turns	$3{,}011$	$64.22$	$64.70$	$0.51$	$+0.48$ $[-0.11,+1.05]$	$47$	$45$
4 turns	$1{,}571$	$74.23$	$74.32$	$0.27$	$+0.09$ $[-0.31,+0.54]$	$41$	$40$
8 turns	$848$	$81.23$	$81.31$	$0.14$	$+0.07$ $[-0.58,+0.62]$	$31$	$31$
session	$272$	$88.91$	$88.95$	$0.05$	$+0.04$ $[-0.28,+0.29]$	$19$	$15$

Appendix

Table D.7: Granularity ladder (BM25 replay, online count rule, Recall@10 over gold units, conversation-cluster

95\%

intervals; recall levels are not comparable across rungs). Purity is the mean share of gold turns in a retrieved gold unit.

Observation	Seed	Rate (%)	$\Delta$ Recall [95% CI]
traffic stats	1	52.4	$+1.05$ $[+0.40,+1.64]$
	2	58.6	$+1.08$ $[+0.17,+1.88]$
	3	71.6	$+1.58$ $[+0.88,+2.22]$
stats + text	1	46.7	$+1.13$ $[+0.49,+1.69]$
	2	49.3	$+1.33$ $[+0.77,+1.92]$
	3	63.3	$+1.50$ $[+0.79,+2.23]$

Appendix

Table E.1: Observation ablation in the matched BM25 replay. Values are recall-point differences with conversation-cluster

95\%

intervals; rate is the fraction of feedback entries suppressed.

Deployment	Host	q / store	RO	Separable unit	Headroom	Outcome
LoCoMo	BM25	154	$0.74$	yes (turns)	$27.5\%$	helps
LoCoMo	Mem0	154	–	yes (facts)	–	helps
LoCoMo	MemoryOS	154	$0.81$	pages ( $\leq 14$ turns)	$22.8\%$	helps
LoCoMo	A-MEM	154	–	roots expand into links	–	all policies improve F1
MemBench, history adapter	BM25	1 final	–	yes	–	count loses; others gain
GroupMem (4 domain stores)	BM25	145–214	–	yes	–	count loses; others gain

Appendix

Table F.1: Deployment diagnostics for persistent index adaptation. Rows combine native-host results with explicitly marked BM25 replay interventions on granularity and store content. Dashes denote unmeasured or undefined quantities.

Policy	Seed	Model evaluations	Median (ms)	P95 (ms)	Counters	State bytes
off	–	0	-	-	0	2
count	–	0	-	-	235	9,669
frozen	–	1,540	23,997	25,400	152	6,255
trained	20260908	1,540	87.3	88.9	190	7,818
	20260909	1,540	86.9	88.2	187	7,712
	20260910	1,540	86.6	88.3	216	8,890

Appendix

Table G.1: Recorded policy cost on native Mem0 ,

1{,}540

queries per arm. “Counters” is the number of memories the state ever touches, and “state bytes” the serialised size of the persistent index state.

System	Native adaptation target	Persistent access adaptation	Separate index-only policy
BM25	Static sparse retrieval	None	NR
Mem0 ( Chhikara et al., 2025 )	Fact consolidation	Updated memory records	NR
MemoryOS ( Kang et al., 2025 )	Tiered memory management	Heat and placement	NR
A-MEM ( Xu et al., 2025 )	Notes and linked metadata	Links coupled to note evolution	NR
Memory-R2 ( Yan et al., 2026a )	Learned memory management	Reasoned insert, update, and delete operations	NR
AMIE	Persistent access weights	Count and semantic suppression	Explicitly isolated

Appendix

Table I.1: Native memory and index adaptation mechanisms in the evaluated hosts. NR: separate index-only policy not reported.

Agent Memory

Agent memory lets AI systems retain, update, and use facts, experiences, and skills across sessions and long-horizon tasks.

See all 192 papers

CoEvo-Mem: Co-Evolving Retrieval Policy and Memory Bank for LLM Agents
Aug 3, 2026Bowen Ye, Yongchao Xu, Zhijian Li +3Continual Learning for LLM Agents Adaptive RAG
Alibaba Group, Hangzhou, China · School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, Shanghai 200240, China · University of Science and Technology of China, Hefei, China
EvolveMem:Self-Evolving Memory Architecture via AutoResearch for LLM Agents
May 13, 2026Jiaqi Liu, Xinyu Ye, Peng Xia +4LLM Agent Memory LLM Memory
UNC-Chapel Hill · UC Berkeley · UCSC
ERSkill: Evolving for Skill-Guided Adaptive Memory Retrieval
Aug 13, 2026Haolong Chen, Liang Zhang, Zhuo Li +2LLM Agent Memory LLM Agent Skill Learning
Shenzhen International Center for Industrial and Applied Mathematics · Shenzhen Research Institute of Big Data · The Chinese University of Hong Kong, Shenzhen +2

Aug 3, 2026J/K move · Enter open · S save

CoEvo-Mem: Co-Evolving Retrieval Policy and Memory Bank for LLM Agents

Bowen Ye, Yongchao Xu, Zhijian Li, Xiang Yin+2

Alibaba Group, Hangzhou, China · School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, Shanghai 200240, China · University of Science and Technology of China, Hefei, China

Continual Learning for LLM Agents Adaptive RAG Memory-Augmented Language Model Agents

Open paper AskPDF arXiv

As memories accumulate across tasks and sessions, the performance of long-term LLM agents depends jointly on query-specific retrieval and continual memory refinement. However, existing methods typically optimize either memory access, through iterative query refinement or adaptive retrieval policies, or memory evolution such as structural update. This separation overlooks a fundamental feedback loop: retrieval determines which memories receive usage signals, while updated memory bank reshape future retrieval. We propose CoEvo-Mem, a closed-loop framework for co-evolving the retrieval policy and memory bank. For each query, a frozen LLM generates route-specific query rewrites and a routing prior, which a lightweight residual router corrects online. The retrieved context serves as the coupling interface between the two learning processes: task outcomes assign credit to routing decisions, while trajectory-conditioned feedback updates memory values and graph relations. These updates alter how memories are ranked and selected for subsequent queries, thereby closing the feedback loop. To mitigate coupling induced non-stationarity, CoEvo-Mem alternates between updating the router with the memory bank fixed and evolving the memory bank with the retrieval policy fixed. Across seven diverse benchmarks, CoEvo-Mem achieves state-of-the-art performance, demonstrating the importance of retrieval-memory coevolution.

Abstract

Agent Memory

Explore similar work

CoEvo-Mem: Co-Evolving Retrieval Policy and Memory Bank for LLM Agents

EvolveMem:Self-Evolving Memory Architecture via AutoResearch for LLM Agents

ERSkill: Evolving for Skill-Guided Adaptive Memory Retrieval