After the Fix: Transfer of Corrected Agent Experience

Yanfei Zhang; Xu Lin

cs.AISep 28, 2026

After the Fix: Transfer of Corrected Agent Experience

Authors: Yanfei Zhang, Xu Lin

Organizations: International Digital Economy Academy (IDEA)

Abstract

Does repairing an episode make its experience a better memory for the next task? We transfer the same failed source before and after accepted repair to a fixed target, alongside independent execution. Our 3,300 runs cover 100 ThinkingBox pairs and the same 100 APEX pairs with and without source-state inheritance, under eleven conditions. ThinkingBox's Full/Skill/Hybrid correction gains are 44/29/32 percentage points, with corrected performance 25/22/18 points above independence; inference weakens at the task-family level. Yet 12 of Full's 15-point larger correction gap over Skill come from worse uncorrected performance, not better corrected memory. Moreover, 22 of Full's 46 upward transitions restore observed baseline success. Neither APEX regime establishes comparable aggregate correction benefits. Action evidence connects workflow gains with reusable obligations and convention conflicts with source-local choices. Text APEX's accepted execution reaches 52% versus its summary's 40%, without robust global/group-level superiority or an estab- lished advantage over independence. Smaller handoffs reduce input but increase calls. The value of repairing experience is therefore distinct from the value of reusing it: memory updates require both a previous-version reference and a fresh-start reference.

Figures & tables

Figure 1: The same failed A branches into U/C; each handoff runs B independently. State and text are separate channels. Table 1 defines all six handoffs, including the benchmark-specific Hybrid constructors. None retains the shared state channel; B-only receives no A experience.

Format	Episode content given to B
None	Shared state channel only; no additional episode text.
Full	Chronological actions, observations, and any failure/feedback/repair, subject to trace limits.
Skill	Abstract procedures, constraints, checks, and error–correction rules.
Hybrid	ThinkingBox: history summary plus four recent raw assistant segments. APEX: the same Skill plus grounded facts and provenance.
Summary	C-only; at most 350 requested words on accepted results, entities, and decisive actions; no earlier failures/feedback.
Execution (Exec)	Identical Summary plus the final accepted segment’s actions, tool calls, and observations; not the entire repair history.

Table 1: A-derived episode content in each handoff. Every format also retains its regime’s shared state channel. No constructor sees B’s request or score.

Condition	ThinkingBox	State	Text
B-only	42	44	47
U-None / C-None	37 / 36	45 / 44	45 / 47
U-Full / C-Full	23 / 67	42 / 41	45 / 45
U-Skill / C-Skill	35 / 64	39 / 42	44 / 45
U-Hybrid / C-Hybrid	28 / 60	41 / 45	48 / 40
Summary / Execution	56 / 67	40 / 44	40 / 52

Table 2: Absolute target acceptance (%). Every cell has 100 targets, including budget failures. State/Text are the two APEX regimes.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Regime	B’s starting environment	A-derived information
ThinkingBox	Independent standard B sandbox	Terminal-state description and the specified textual representation; no live A database restoration
APEX-state	$W\rightarrow I_{A}\rightarrow\Delta_{A}^{s}\rightarrow I_{B}$	Source environment state plus the specified text; None supplies the state channel alone
APEX-text	$W\rightarrow I_{B}$	Text alone; None supplies terminal observations, with the same branch prefix in Full/Skill/Hybrid

Appendix

Table 3: State and text channels.

W

is the original APEX world,

I_{A}/I_{B}

the task inputs, and

\Delta_{A}^{s}

branch-specific source changes. B-only excludes A in every regime.

Contrast	Held fixed	What the contrast measures
C–U, same representation	Target and evaluation	The complete correction-associated experience package
C–B-only	Target and evaluation	Benefit beyond independent target execution
Execution–Summary	Exact summary, framing, starting state	Adding the accepted segment, including its content and length
Full–Execution	Target and evaluation	Alternative whole handoffs, not history alone
APEX-text–APEX-state	Pair identities and target rubric	Implemented handoff regimes, with state and textual-construction differences

Appendix

Table 4: Experimental contrasts and their interpretation.

Component	ThinkingBox	APEX-state
Acting model	Environment-resolved DeepSeek alias; adapter default deepseek-v4-flash	DeepSeek V4.1 Flash (multimodal), API deepseek-flash
Agent sampling	Thinking enabled; explicit reasoning_effort=high ; temperature omitted	Thinking enabled; no explicit reasoning effort; temperature 0.2; no supplied seed
Agent budgets	32,768 output; 400 interactions; 30 user turns; 1,200 seconds	8,192 output; 100 steps; 7,200 seconds
Runtime user	Temperature 0.2; 4,096 output; cannot terminate agent conversation	No separate conversational user during B
Correction review	Temperature 0.2; 2,048 output; trace and state differences	Temperature 0; 8,192 output; artifacts and criteria
Memory compressor	Temperature 0; 8,192 output; thinking disabled	deepseek-v4-pro ; temperature 0; 8,192 output

Appendix

Table 5: Execution settings. Output budgets are tokens per call. The package contains full prompt constructors and saved configuration fields.

Condition	ThinkingBox	APEX-state A/Q	APEX-text A/Q
B-only	42	44 / 55.54	47 / 57.40
U-None	37	45 / 54.79	45 / 53.50
U-Full	23	42 / 54.00	45 / 55.83
U-Skill	35	39 / 50.89	44 / 54.29
U-Hybrid	28	41 / 53.11	48 / 56.80
C-None	36	44 / 55.38	47 / 56.10

Appendix

Table 6: All eleven conditions in all three regimes. Each entry has denominator 100. A/Q denotes acceptance percentage / task-normalized rubric percentage.

Regime	Cmp.	G/L	Gap	Cluster CI	Raw $p$	$p_{H}$	$p_{24}$
TB	C-F / U-F	46/2	+44	[25.3, 56.3]	$<0.0001$	$<0.0001$	$<0.0001$
TB	C-S / U-S	39/10	+29	[12.7, 39.8]	$<0.0001$	0.0003	0.0008
TB	C-H / U-H	36/4	+32	[19.1, 44.0]	$<0.0001$	$<0.0001$	$<0.0001$
TB	C-F / B	31/6	+25	[13.1, 37.7]	$<0.0001$	0.0004	0.0009
TB	C-S / B	30/8	+22	[8.0, 39.0]	0.0005	0.0038	0.0094
TB	C-H / B	24/6	+18	[6.7, 32.5]	0.0014	0.0100	0.0272

Appendix

Table 7: All 24 paired acceptance contrasts. G/L counts gains/losses; differences and world/family-cluster 95% intervals are percentage points. TB/State/Text denote the three regimes; F/S/H in condition names denote Full/Skill/Hybrid, and standalone B/S/E denote B-only/Summary/Execution.

p_{H}

uses the relevant 9/9/6 family;

p_{24}

uses all contrasts.

Regime	Hypothesis	Gap	Cluster CI	Group $p$	$p_{H}$	$p_{24}$
TB	H1a-full	+44	[25.3, 56.2]	0.0054	0.0439	0.1235
TB	H1a-skill	+29	[12.8, 40.0]	0.0259	0.1812	0.4917
TB	H1a-hybrid	+32	[19.3, 44.0]	0.0049	0.0439	0.1172
TB	H1b-full	+25	[12.9, 37.6]	0.0078	0.0703	0.1719
TB	H1b-skill	+22	[8.0, 39.0]	0.0203	0.1418	0.4053
TB	H1b-hybrid	+18	[6.5, 32.5]	0.0137	0.1094	0.2871

Appendix

Table 8: Primary contrasts under group-level inference. CI: freshly recomputed pointwise cluster interval (pp);

p_{H}

uses the 9/9/6 families,

p_{24}

all primary contrasts.

Regime	Comparison	G/L	Gap	Cluster CI	Target $p_{H}$	Group $p_{H}$
TB	U-F–U-N	4/18	-14	[-27.2, -4.3]	0.1433	0.7734
TB	U-S–U-N	12/14	-2	[-10.8, 4.8]	1.0000	1.0000
TB	U-H–U-N	9/18	-9	[-25.4, 3.3]	1.0000	1.0000
TB	U-S–U-F	16/4	+12	[2.0, 22.9]	0.3782	1.0000
TB	U-H–U-F	10/5	+5	[-4.3, 12.9]	1.0000	1.0000
TB	U-H–U-S	8/15	-7	[-20.7, 3.6]	1.0000	1.0000

Appendix

Table 9: Within-branch comparisons. N/F/S/H: None/Full/Skill/Hybrid. G/L denotes observed upward/downward transitions. The last columns apply Holm across all 36 tests.

Regime	Format contrast	Gap difference	Cluster CI	Group $p_{H}$
TB	Full–None	+45	[27.8, 61.9]	0.0352
TB	Skill–None	+30	[11.2, 48.6]	0.3945
TB	Hybrid–None	+33	[14.9, 53.9]	0.0623
TB	Skill–Full	-15	[-28.7, -1.4]	1.0000
TB	Hybrid–Full	-12	[-22.6, 2.4]	1.0000
TB	Hybrid–Skill	+3	[-11.3, 21.4]	1.0000

Appendix

Table 10: Difference in correction gaps between representations. All intervals are pp; group

p_{H}

adjusts the 18 interactions.

	ThinkingBox			APEX-state			APEX-text
B/U/C	F	S	H	F	S	H	F	S	H
000	26	22	31	43	40	42	38	37	38
001	24	21	16	4	7	6	5	5	2
010	1	6	3	3	4	3	4	7	8
011	7	9	8	6	5	5	6	4	5
100	5	4	5	7	6	5	7	5	9

Appendix

Table 11: All eight outcome patterns, including unchanged outcomes. Each column sums to 100. F/S/H denote Full/Skill/Hybrid.

Relationship	Pairs	Shared basis	What changes in B
Same-family variant	42	Same computation, model, or legal decision test	LBO debt/rate assumptions; a different insured claim
Shared intermediate procedure	26	Identifiable reusable subprocedure; distinct end operation	Comparable filtering for peer benchmarking versus LBO inputs
Shared evidence	20	Common case materials or model; distinct operations	Manual drafting versus compliance comparison
Context only	12	Same project or domain; no specific shared procedure or material established	Board-tenure comparison versus CAPM estimation

Appendix

Table 12: Relationship definitions and examples from the request-based audit.

Relationship	Count	Full	Skill	Hybrid	E–S
Same-family	42	-4.8 / +0.0	+7.1 / -7.1	-2.4 / -9.5	+2.4 / +14.3
Shared procedure	26	+0.0 / +0.0	+0.0 / +19.2	+19.2 / -3.8	+3.8 / +3.8
Shared evidence	20	-5.0 / +5.0	-5.0 / -5.0	-5.0 / -5.0	+15.0 / +15.0
Context only	12	+16.7 / -8.3	+8.3 / +0.0	+8.3 / -16.7	-8.3 / +16.7

Appendix

Table 13: Matched APEX relationship strata. Every effect is shown as state / text in percentage points. Full/Skill/Hybrid are C–U; E–S is Execution–Summary.

Domain	Count	Full	Skill	Hybrid	E–S
Banking	48	-10.4 / -6.2	+0.0 / -6.2	+0.0 / -18.8	+14.6 / +10.4
Law	23	+4.3 / +0.0	+17.4 / +30.4	+17.4 / +4.3	-8.7 / +21.7
Consulting	29	+10.3 / +10.3	-3.4 / -10.3	+0.0 / +0.0	-3.4 / +6.9

Appendix

Table 14: Domain-specific contrasts, again state / text in percentage points. Same definitions as Table 13 ; these are exploratory subgroup results.

	Improved			Regressed
Regime	$\geq 1$	$\geq 2$	All 3	$\geq 1$	$\geq 2$	All 3
ThinkingBox	61	42	18	15	1	0
APEX-state	29	5	1	22	6	1
APEX-text	22	6	1	28	7	1

Appendix

Table 15: Targets improved or regressed under at least one, at least two, or all three C–U representations. Improvement and regression columns are not disjoint: a target can improve under one representation and regress under another.

Contrast	State gains	Text gains	Gain in both	Loss in both	G to L	L to G
Full	8	10	2	2	0	0
Skill	15	14	5	5	3	2
Hybrid	12	5	2	4	4	0
Execution–Summary	12	14	2	0	0	4

Appendix

Table 16: Matching the same target’s contrast across APEX regimes. Gains/losses mean C–U, except E–S. “G to L” means a state gain becomes a text loss; “L to G” is the reverse. Each row has 100 matched pairs.

Contrast	State	Text	Difference	World interval
C-Full–U-Full	-1	0	+1	$[-6.19,6.67]$
C-Skill–U-Skill	+3	+1	-2	$[-15.74,12.77]$
C-Hybrid–U-Hybrid	+4	-8	-12	$[-24.00,1.10]$
Execution–Summary	+4	+12	+8	$[-5.46,24.27]$
Full–Execution	-3	-7	-4	$[-19.80,7.56]$

Appendix

Table 17: Within-regime effects and matched differences of effects. The last column is a world-cluster 95% interval for text minus state, in percentage points.

Comparison	Before	After	Removed	Then pass	Still fail
U-Full to C-Full	55	2	54	37	17
U-Skill to C-Skill	40	7	39	28	11
U-Hybrid to C-Hybrid	48	3	48	30	18
Summary to Execution	13	1	13	8	5

Appendix

Table 18: Recorded lifecycle violations, all 100 paired targets in each comparison. “Removed” counts reference-flagged targets without the flag in the candidate. Its last two columns partition those targets by candidate acceptance.

Regime	Comparison	Clusters	Minimum	Maximum
ThinkingBox	C-F / U-F	15	+40.66	+48.42
ThinkingBox	C-S / U-S	15	+25.27	+31.96
ThinkingBox	C-H / U-H	15	+27.78	+34.74
ThinkingBox	E / S	15	+8.64	+13.33
APEX-state	C-F / U-F	16	-2.22	+1.08
APEX-state	C-S / U-S	16	+1.05	+4.30

Appendix

Table 19: Leave-one-cluster-out acceptance gaps (pp), including all three regimes. Each row removes one complete family or world at a time. Min/max are influence ranges, not confidence intervals. Abbreviations follow Table 7 .

Requested quantity	Summary	Execution	Reference
Current stake	4,935.9	5,499.7	5,499.7
PV of revised cash flows	4,389.4	4,790.1	4,790.1
Discounted terminal value	17,822.9	19,959.1	19,959.0
Revised stake	4,442.5	4,949.8	4,949.8
Percentage loss	10.0%	10.0%	10.0%
Criteria satisfied	1/5	5/5	–

Appendix

Table 20: 3M downside analysis. Dollar values are millions. All conditions produce a deck; model lineage changes the numerical findings.

Regime	Summary	Execution	Full	G/L	E–S	$p_{H}$
ThinkingBox	56	67	67	15/4	+11	0.0961
APEX-state	40	44	41	12/8	+4	1.0000
APEX-text	40	52	45	14/2	+12	0.0251

Appendix

Table 21: Accepted-experience conditions. G/L counts paired gains/losses for Execution–Summary;

p_{H}

is the six-test H2-family adjustment.

Regime	Condition	Accepted	Input M	Calls	Input/call k	Tools
ThinkingBox	B-only	42	0.232	9.48	24.4	14.96
ThinkingBox	Summary	56	0.264	9.48	27.9	17.09
ThinkingBox	Execution	67	0.353	9.27	38.1	18.34
ThinkingBox	C-Full	67	0.434	8.85	49.0	17.07
ThinkingBox	C-Skill	64	0.286	9.80	29.2	17.61
ThinkingBox	C-Hybrid	60	0.318	9.82	32.4	17.44

Appendix

Table 22: B-stage resources for B-only and corrected handoffs. Each mean includes all 100 targets, including failures. Input is cumulative millions of tokens; input/call is total input divided by total calls in thousands.

Regime / set	Completed	Budget failed	Error	Total
State: original nine conditions	895	5	0	900
State: Summary	92	8	0	100
State: Execution	88	10	2	100
Text: all eleven conditions	1,072	28	0	1,100

Appendix

Table 23: APEX execution coverage. Completed means agent submission, not owner acceptance. Budget and infrastructure failures remain distinct. Every cell has a judgment in the aligned analysis.

Study	How experience transfers	Main comparison axis
ExpeL ( Zhao et al., 2024 )	Knowledge extraction and experience retrieval	Accumulation and transfer
CLIN ( Majumder et al., 2024 )	Causal abstractions and meta-memory	Adaptation and generalization
AWM ( Wang et al., 2025 )	Induced reusable workflows	Offline/online and cross-domain
ReasoningBank ( Ouyang et al., 2026 )	Strategies from success and failure	Memory; exploration via MaTTS
Feng et al. (2026)	Task- or subtask-induced skills	Granularity $\times$ text/code
SkillsBench ( Li et al., 2026a )	Curated or generated skills	Skill source, domain, and utility

Appendix

Table 24: Comparison units and experimental axes in cross-task experience reuse.

Explore similar work

Causal Episodic Memory for Feedback-Driven Agent Repair
Date pendingKhang Nhat Hoang Vo, Tam Minh Chu, Anh Trac Duc Dinh +2Repair Episodic Memory
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates · Faculty of Computer Science and Engineering, Ho Chi Minh City University of Technology (HCMUT), VNU-HCM, Ho Chi Minh City, Vietnam
From Faulty Memories to Corrected Actions: Dependency-Guided Rollback Repair for Memory-Augmented Agents
Aug 11, 2026Caili Yu, Yiqi Wang, Jiaqi Zhang +5Rollback Stress Testing
MEMOREPAIR: Barrier-First Cascade Repair in Agentic Memory
May 8, 2026Yang Zhao, Chengxiao Dai, Mengying Kou +1Agentic Memory Nativemem

Date pending, cs.CLJ/K move · Enter open · S save

Causal Episodic Memory for Feedback-Driven Agent Repair

Khang Nhat Hoang Vo, Tam Minh Chu, Anh Trac Duc Dinh, Thuyen Vinh Ha Bui+1

Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates · Faculty of Computer Science and Engineering, Ho Chi Minh City University of Technology (HCMUT), VNU-HCM, Ho Chi Minh City, Vietnam

Repair Episodic Memory Oracles Text-To-Sql

Open paper AskPDF arXiv

LLM agents that repair failures often discard successful corrections, forcing later episodes to rediscover similar solutions. We study whether finalized repair outcomes can improve subsequent Text-to-SQL episodes without parameter updates. We introduce MERIT, a training-free agent that maintains an online dual-polarity memory of oracle-verified corrections and observed unsuccessful directions. Under oracle-assisted benchmark feedback, only memories from earlier finalized episodes are eligible for retrieval. A deterministic classifier assigns a coarse failure type, which conditions a hybrid lexical-dense retriever before the frozen model generates each revision. Using Qwen2.5-7B-Instruct with identical initial predictions and repair budgets, \method{} improves execution accuracy over stateless iterative repair from

66.34\%

69.79\%

on Spider and from

47.35\%

48.44\%

on BIRD. Paired analyses provide clear evidence for the Spider gain but weaker evidence on BIRD. MERIT is not reliably separated from untyped dynamic retrieval on either benchmark, while Reflexion-style memory reaches

51.24\%

on BIRD at substantially higher inference cost. Ablations show that negative memory contributes modestly, the value of type conditioning and lexical-dense ranking is dataset dependent, and schema-local experience provides the most consistent benefit. These results clarify when causal cross-query memory improves repair and when broader memory representations remain preferable. Our implementation is available here:

Abstract

Explore similar work

Causal Episodic Memory for Feedback-Driven Agent Repair

From Faulty Memories to Corrected Actions: Dependency-Guided Rollback Repair for Memory-Augmented Agents

MEMOREPAIR: Barrier-First Cascade Repair in Agentic Memory