Reliable Knowledge Acquisition for Continual Learning Agents

Yixu Huang; Jiaqi W. Ma

cs.CLOct 15, 2026

Reliable Knowledge Acquisition for Continual Learning Agents

Authors: Yixu Huang, Jiaqi W. Ma

Organizations: University of California, Davis · University of Illinois Urbana-Champaign

Abstract

As agents powered by foundation models move into open-ended environments, continual learning from experience becomes essential. For agents built on models with fixed parameters, this learning requires reusing prior task experience in context to improve performance and efficiency on future tasks. Existing memory and experience-learning systems typically focus on either retained facts or procedures, and lack a mechanism for establishing the reliability of claims extracted from the ongoing task stream. Indeed, we find that approximately 12--33% of entries produced by representative systems harm future task performance when reused. To address these limitations, we propose RIKA (Reliable Knowledge Acquisition), which combines hierarchical extraction of factual, conceptual, procedural, and metacognitive knowledge with Evidence-based Knowledge Admission, which uses source evidence and subsequent experience to govern which claims enter the task agent's context, without additional validation tasks. Across six task streams from Continual Learning Bench and four model backbones, RIKA achieves the highest aggregate score on every backbone.

Figures & tables

Figure 1: Extracted knowledge that fails under reuse.

Method	DB	Cohort	Codebase
Mem0	11.7%	13.4%	12.5%
ACE	20.2%	18.9%	23.1%
Trace2Skill	32.7%	28.8%	26.0%
SkillOpt	25.9%	26.1%	28.0%

Table 1: Harmful entry rates under leave-one-entry-out evaluation. Rates are computed over entries within each artifact and averaged equally across artifacts.

Method	DB	Cohort	Codebase
SkillOpt	0.147	-0.062	0.567
SkillOpt + EKA	0.192	-0.045	0.573
Trace2Skill	0.158	-0.053	0.538
Trace2Skill + EKA	0.224	-0.043	0.570

Table 2: Task performance of SkillOpt and Trace2Skill with and without EKA. Scores are averaged over five runs. Each + EKA variant retains the system’s extractor and applies EKA thereafter.

Figure 2: RIKA: hierarchical extraction with evidence-based knowledge admission.

Model	Method	DB		Poker		Cohort		BSM		Sales		Codebase		Agg.
Model	Method	Score	Gain	Score	Gain	Score	Gain	Score	Gain	Score	Gain	Score	Gain	Score	Gain
GPT-5.4	RIKA	0.351	+0.213	1.247	+0.134	-0.020	-0.013	0.579	+0.360	0.808	+0.342	0.551	+0.053	0.232	+0.232
	ICL	0.347	+0.209	0.798	-0.315	-0.028	-0.021	0.513	+0.294	0.769	+0.303	0.534	+0.036	0.183	+0.183
	Mem0	0.431	+0.293	0.612	-0.501	-0.034	-0.027	0.375	+0.156	0.651	+0.185	0.584	+0.086	0.140	+0.140
	ACE	0.196	+0.058	1.196	+0.083	-0.004	+0.003	0.220	+0.001	0.510	+0.044	0.609	+0.111	0.067	+0.067
	Trace2Skill	0.138	+0.000	1.115	+0.002	-0.032	-0.025	0.318	+0.099	0.689	+0.223	0.513	+0.015	0.071	+0.071

Table 3: Continual-learning performance across streams and backbones. Score : stream-level score; Gain : improvement over the shared stateless control for the same backbone and task; Agg. : averages of normalized scores and gains across six streams. See Appendices E and F for setup details and 95% confidence intervals.

Panel of Figure 3 — Figure 3: Cumulative gain on DB and Cohort with DeepSeek-V4-Flash. Curves show mean gain relative to the stateless control shared across methods within each task; shading indicates 95% confidence intervals. The vertical line marks the DB migration at instance 20.

Method	DB	Cohort	Codebase
RIKA	0.281	-0.042	0.575
Imm. Only	0.225	-0.089	0.540
Def. Only	0.265	-0.070	0.557
Random Router	0.249	-0.084	0.549

Table 4: Admission-routing ablations.

Method	DB	Cohort	Codebase
RIKA	0.281	-0.042	0.575
Generic 1-type	0.213	-0.087	0.541
Flat 4-type	0.261	-0.057	0.563
$-$ Factual	0.253	-0.060	0.547
$-$ Conceptual	0.265	-0.065	0.553
$-$ Procedural	0.270	-0.073	0.549

Table 5: Ablations of knowledge types and hierarchical extraction.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Method	DB	Cohort	Codebase
RIKA	3.7	6.2	5.7
Immediate-only	22.1	16.3	20.1
Deferred-only	5.9	5.3	6.9
Random Router	14.2	19.6	15.2

Appendix

Table 6: Harmful-entry rates (%) under admission-routing ablations. All variants follow the same harmful-entry evaluation protocol.

Relation or event	Claim and graph update	Evidence and admission update
Equivalent	Merge claims and rewire edges.	Incorporate batch support counts; conflicting routes fall back to deferred.
Specialize	Keep both; add $u\rightarrow v$ .	Transfer support along $u\rightarrow v$ and onward to reachable generalizations.
Generalize	Keep both; add $v\rightarrow u$ .	Transfer support along $v\rightarrow u$ and onward to reachable generalizations.
Contradict	Keep both; annotate conflict in $H_{j}$ .	Apply the batch opposition count to the target; assess both claims by their own thresholds.
Overlap	Keep both; annotate overlap in $H_{j}$ .	Preserve each claim’s evidence count.
Independent	Keep claims separate and preserve their existing relations.	Initialize unmatched claims from their batch evidence counts; leave existing evidence unchanged.

Appendix

Table 7: State updates for a new claim

u

and stored claim

v

. Evidence updates precede forgetting and batch-wise admission.

Backbone	DB	Poker	Cohort	BSM	Sales	Codebase
GPT-5.4	0.138	1.113	-0.007	0.219	0.466	0.498
DeepSeek-V4-Flash	0.060	0.652	-0.118	0.219	0.514	0.512
MiniMax-M3	0.136	-0.135	-0.005	0.220	0.197	0.373
Gemini-3-Flash	0.089	1.640	-0.059	0.219	0.439	0.407

Appendix

Table 8: Shared stateless performance averaged over five runs. For each backbone and task stream, the reference is the mean of the five stateless controls from the corresponding ICL evaluation. All methods in Table 3 use this same reference to compute Gain and normalized aggregate Gain; the ablation and sensitivity analyses use the corresponding DeepSeek-V4-Flash reference for Gain.

Method	Metric	DB	Poker	Cohort	BSM	Sales	Codebase	Agg.
RIKA	Score	$0.351\,\pm\,0.019$	$\mathbf{1.247}\,\pm\,0.097$	$-0.020\,\pm\,0.007$	$\mathbf{0.579}\,\pm\,0.035$	$\mathbf{0.808}\,\pm\,0.023$	$0.551\,\pm\,0.018$	$\mathbf{0.232}\,\pm\,0.029$
	Gain	$+0.213\,\pm\,0.019$	$\mathbf{+0.134}\,\pm\,0.097$	$-0.013\,\pm\,0.007$	$\mathbf{+0.360}\,\pm\,0.035$	$\mathbf{+0.342}\,\pm\,0.023$	$+0.053\,\pm\,0.018$	$\mathbf{+0.232}\,\pm\,0.029$
ICL	Score	$0.347\,\pm\,0.018$	$0.798\,\pm\,0.152$	$-0.028\,\pm\,0.006$	$0.513\,\pm\,0.023$	$0.769\,\pm\,0.014$	$0.534\,\pm\,0.009$	$0.183\,\pm\,0.021$
	Gain	$+0.209\,\pm\,0.018$	$-0.315\,\pm\,0.152$	$-0.021\,\pm\,0.006$	$+0.294\,\pm\,0.023$	$+0.303\,\pm\,0.014$	$+0.036\,\pm\,0.009$	$+0.183\,\pm\,0.021$
Mem0	Score	$\mathbf{0.431}\,\pm\,0.022$	$0.612\,\pm\,0.044$	$-0.034\,\pm\,0.006$	$0.375\,\pm\,0.034$	$0.651\,\pm\,0.033$	$0.584\,\pm\,0.025$	$0.140\,\pm\,0.027$
	Gain	$\mathbf{+0.293}\,\pm\,0.022$	$-0.501\,\pm\,0.044$	$-0.027\,\pm\,0.006$	$+0.156\,\pm\,0.034$	$+0.185\,\pm\,0.033$	$+0.086\,\pm\,0.025$	$+0.140\,\pm\,0.027$

Appendix

Table 9: GPT-5.4 results with 95% confidence intervals. Means match Table 3 . Task-level Gain intervals condition on the fixed shared control.

Method	Metric	DB	Poker	Cohort	BSM	Sales	Codebase	Agg.
RIKA	Score	$\mathbf{0.281}\,\pm\,0.016$	$\mathbf{1.141}\,\pm\,0.121$	$-0.042\,\pm\,0.016$	$0.234\,\pm\,0.039$	$\mathbf{0.847}\,\pm\,0.039$	$\mathbf{0.575}\,\pm\,0.020$	$\mathbf{0.141}\,\pm\,0.031$
	Gain	$\mathbf{+0.221}\,\pm\,0.016$	$\mathbf{+0.489}\,\pm\,0.121$	$+0.076\,\pm\,0.016$	$+0.015\,\pm\,0.039$	$\mathbf{+0.333}\,\pm\,0.039$	$\mathbf{+0.063}\,\pm\,0.020$	$\mathbf{+0.233}\,\pm\,0.031$
ICL	Score	$0.256\,\pm\,0.010$	$0.755\,\pm\,0.098$	$-0.084\,\pm\,0.012$	$\mathbf{0.518}\,\pm\,0.033$	$0.639\,\pm\,0.042$	$0.548\,\pm\,0.016$	$0.074\,\pm\,0.022$
	Gain	$+0.196\,\pm\,0.010$	$+0.103\,\pm\,0.098$	$+0.034\,\pm\,0.012$	$\mathbf{+0.299}\,\pm\,0.033$	$+0.125\,\pm\,0.042$	$+0.036\,\pm\,0.016$	$+0.176\,\pm\,0.022$
Mem0	Score	$0.274\,\pm\,0.022$	$0.732\,\pm\,0.133$	$\mathbf{-0.038}\,\pm\,0.017$	$0.348\,\pm\,0.057$	$0.522\,\pm\,0.051$	$0.547\,\pm\,0.014$	$0.049\,\pm\,0.033$
	Gain	$+0.214\,\pm\,0.022$	$+0.080\,\pm\,0.133$	$\mathbf{+0.080}\,\pm\,0.017$	$+0.129\,\pm\,0.057$	$+0.008\,\pm\,0.051$	$+0.035\,\pm\,0.014$	$+0.129\,\pm\,0.033$

Appendix

Table 10: DeepSeek-V4-Flash results with 95% confidence intervals. Means match Table 3 . Task-level Gain intervals condition on the fixed shared control.

Method	Metric	DB	Poker	Cohort	BSM	Sales	Codebase	Agg.
RIKA	Score	$\mathbf{0.327}\,\pm\,0.013$	$\mathbf{0.623}\,\pm\,0.099$	$\mathbf{0.012}\,\pm\,0.009$	$0.257\,\pm\,0.042$	$0.758\,\pm\,0.079$	$\mathbf{0.554}\,\pm\,0.012$	$\mathbf{0.163}\,\pm\,0.015$
	Gain	$\mathbf{+0.191}\,\pm\,0.013$	$\mathbf{+0.758}\,\pm\,0.099$	$\mathbf{+0.017}\,\pm\,0.009$	$+0.037\,\pm\,0.042$	$+0.561\,\pm\,0.079$	$\mathbf{+0.181}\,\pm\,0.012$	$\mathbf{+0.239}\,\pm\,0.015$
ICL	Score	$0.284\,\pm\,0.019$	$0.407\,\pm\,0.103$	$-0.020\,\pm\,0.007$	$\mathbf{0.337}\,\pm\,0.038$	$0.535\,\pm\,0.055$	$0.483\,\pm\,0.026$	$0.043\,\pm\,0.012$
	Gain	$+0.148\,\pm\,0.019$	$+0.542\,\pm\,0.103$	$-0.015\,\pm\,0.007$	$\mathbf{+0.117}\,\pm\,0.038$	$+0.338\,\pm\,0.055$	$+0.110\,\pm\,0.026$	$+0.147\,\pm\,0.012$
Mem0	Score	$0.303\,\pm\,0.010$	$0.235\,\pm\,0.177$	$0.003\,\pm\,0.005$	$0.289\,\pm\,0.058$	$0.501\,\pm\,0.056$	$0.402\,\pm\,0.027$	$0.018\,\pm\,0.014$
	Gain	$+0.167\,\pm\,0.010$	$+0.370\,\pm\,0.177$	$+0.008\,\pm\,0.005$	$+0.069\,\pm\,0.058$	$+0.304\,\pm\,0.056$	$+0.029\,\pm\,0.027$	$+0.132\,\pm\,0.014$

Appendix

Table 11: MiniMax-M3 results with 95% confidence intervals. Means match Table 3 . Task-level Gain intervals condition on the fixed shared control.

Method	Metric	DB	Poker	Cohort	BSM	Sales	Codebase	Agg.
RIKA	Score	$\mathbf{0.387}\,\pm\,0.012$	$\mathbf{0.952}\,\pm\,0.076$	$-0.058\,\pm\,0.010$	$\mathbf{0.434}\,\pm\,0.033$	$0.628\,\pm\,0.040$	$0.614\,\pm\,0.016$	$\mathbf{0.130}\,\pm\,0.022$
	Gain	$\mathbf{+0.298}\,\pm\,0.012$	$\mathbf{-0.688}\,\pm\,0.076$	$+0.001\,\pm\,0.010$	$\mathbf{+0.215}\,\pm\,0.033$	$+0.189\,\pm\,0.040$	$+0.207\,\pm\,0.016$	$\mathbf{+0.201}\,\pm\,0.022$
ICL	Score	$0.376\,\pm\,0.008$	$0.790\,\pm\,0.069$	$\mathbf{-0.034}\,\pm\,0.004$	$0.367\,\pm\,0.025$	$\mathbf{0.707}\,\pm\,0.035$	$0.390\,\pm\,0.023$	$0.084\,\pm\,0.015$
	Gain	$+0.287\,\pm\,0.008$	$-0.850\,\pm\,0.069$	$\mathbf{+0.025}\,\pm\,0.004$	$+0.148\,\pm\,0.025$	$\mathbf{+0.268}\,\pm\,0.035$	$-0.017\,\pm\,0.023$	$+0.160\,\pm\,0.015$
Mem0	Score	$0.355\,\pm\,0.015$	$0.863\,\pm\,0.123$	$-0.038\,\pm\,0.009$	$0.334\,\pm\,0.050$	$0.544\,\pm\,0.038$	$0.583\,\pm\,0.010$	$0.084\,\pm\,0.022$
	Gain	$+0.266\,\pm\,0.015$	$-0.777\,\pm\,0.123$	$+0.021\,\pm\,0.009$	$+0.115\,\pm\,0.050$	$+0.105\,\pm\,0.038$	$+0.176\,\pm\,0.010$	$+0.153\,\pm\,0.022$

Appendix

Table 12: Gemini-3-Flash results with 95% confidence intervals. Means match Table 3 . Task-level Gain intervals condition on the fixed shared control.

Method	DB		Cohort		Codebase
Method	Score	Gain	Score	Gain	Score	Gain
RIKA	0.281	+0.221	-0.042	+0.076	0.575	+0.063
Imm. Only	0.225	+0.165	-0.089	+0.029	0.540	+0.028
Def. Only	0.265	+0.205	-0.070	+0.048	0.557	+0.045
Random Router	0.249	+0.189	-0.084	+0.034	0.549	+0.037

Appendix

Table 13: Full admission-routing ablation results. The score-only version appears in Table 4 .

Method	DB		Cohort		Codebase
Method	Score	Gain	Score	Gain	Score	Gain
RIKA	0.281	+0.221	-0.042	+0.076	0.575	+0.063
Generic 1-type	0.213	+0.153	-0.087	+0.031	0.541	+0.029
Flat 4-type	0.261	+0.201	-0.057	+0.061	0.563	+0.051
$-$ Factual	0.253	+0.193	-0.060	+0.058	0.547	+0.035
$-$ Conceptual	0.265	+0.205	-0.065	+0.053	0.553	+0.041

Appendix

Table 14: Full ablations of knowledge types and hierarchical extraction. The score-only version appears in Table 5 .

Method	DB	Cohort	Codebase
RIKA	$\mathbf{0.281}\,\pm\,0.016$	$\mathbf{-0.042}\,\pm\,0.016$	$\mathbf{0.575}\,\pm\,0.020$
Imm. Only	$0.225\,\pm\,0.019$	$-0.089\,\pm\,0.014$	$0.540\,\pm\,0.016$
Def. Only	$0.265\,\pm\,0.011$	$-0.070\,\pm\,0.010$	$0.557\,\pm\,0.021$
Random Router	$0.249\,\pm\,0.024$	$-0.084\,\pm\,0.019$	$0.549\,\pm\,0.027$

Appendix

Table 15: Admission-routing ablations with 95% confidence intervals. Entries report mean Score

\pm

95% confidence-interval half-width.

Method	DB	Cohort	Codebase
RIKA	$\mathbf{0.281}\,\pm\,0.016$	$\mathbf{-0.042}\,\pm\,0.016$	$\mathbf{0.575}\,\pm\,0.020$
Generic 1-type	$0.213\,\pm\,0.021$	$-0.087\,\pm\,0.019$	$0.541\,\pm\,0.023$
Flat 4-type	$0.261\,\pm\,0.010$	$-0.057\,\pm\,0.013$	$0.563\,\pm\,0.027$
$-$ Factual	$0.253\,\pm\,0.012$	$-0.060\,\pm\,0.008$	$0.547\,\pm\,0.019$
$-$ Conceptual	$0.265\,\pm\,0.012$	$-0.065\,\pm\,0.007$	$0.553\,\pm\,0.014$
$-$ Procedural	$0.270\,\pm\,0.017$	$-0.073\,\pm\,0.012$	$0.549\,\pm\,0.012$

Appendix

Table 16: Ablations of knowledge types and hierarchical extraction with 95% confidence intervals. Entries report mean Score

\pm

95% confidence-interval half-width.

Method	DB	Cohort	Codebase
RIKA	$\mathbf{0.281}\,\pm\,0.016$	$\mathbf{-0.042}\,\pm\,0.016$	$\mathbf{0.575}\,\pm\,0.020$
$-$ Atomize	$0.267\,\pm\,0.013$	$-0.059\,\pm\,0.010$	$0.564\,\pm\,0.012$
$-$ Forget	$0.263\,\pm\,0.016$	$-0.053\,\pm\,0.015$	$0.568\,\pm\,0.020$
Incremental write	$0.249\,\pm\,0.023$	$-0.060\,\pm\,0.020$	$0.567\,\pm\,0.027$

Appendix

Table 17: Ablations of atomization, forgetting, and state-based rendering. Entries report mean Score

\pm

95% confidence-interval half-width.

Batch Size	DB	Cohort	Codebase
$n=1$	$0.257\,\pm\,0.018$	$-0.058\,\pm\,0.009$	$0.567\,\pm\,0.017$
$n=3$	$\mathbf{0.285}\,\pm\,0.021$	$\mathbf{-0.039}\,\pm\,0.018$	$0.574\,\pm\,0.014$
$n=5$	$0.281\,\pm\,0.016$	$-0.042\,\pm\,0.016$	$\mathbf{0.575}\,\pm\,0.020$
$n=10$	$0.277\,\pm\,0.016$	$-0.046\,\pm\,0.013$	$0.573\,\pm\,0.016$

Appendix

Table 18: Sensitivity to batch size

n

. Entries report mean Score

\pm

95% confidence-interval half-width.

Threshold	DB	Cohort	Codebase
$k=2$	$0.267\,\pm\,0.008$	$-0.046\,\pm\,0.015$	$0.574\,\pm\,0.015$
$k=3$	$\mathbf{0.281}\,\pm\,0.016$	$-0.042\,\pm\,0.016$	$0.575\,\pm\,0.020$
$k=4$	$0.279\,\pm\,0.013$	$\mathbf{-0.038}\,\pm\,0.011$	$\mathbf{0.576}\,\pm\,0.029$
$k=5$	$0.278\,\pm\,0.021$	$-0.040\,\pm\,0.007$	$0.575\,\pm\,0.023$

Appendix

Table 19: Sensitivity to admission threshold

k

. Entries report mean Score

\pm

95% confidence-interval half-width.

System	GPT-5.4	DeepSeek V4 Flash	Gemini 3 Flash	MiniMax-M3
ICL	$16.32	$0.75	$3.22	$1.60
Mem0	$16.43	$0.64	$3.35	$1.57
ACE	$62.35	$2.33	$11.98	$6.44
Trace2Skill	$23.74	$1.03	$5.21	$2.35
SkillOpt	$46.33	$2.15	$8.33	$4.57
RIKA	$25.89	$1.21	$5.11	$2.53

Appendix

Table 20: Costs corresponding to the main results. Costs are reported in USD for the systems and model backbones evaluated in Table 3 .

Continual Learning for LLM Agents

Continual learning for LLM agents studies how agents adapt from experience while retaining skills and knowledge across tasks.

See all 196 papers

Aug 19, 2026

Harness Continual Learning: Continual Adaptation Beyond Model Parameters

Continual learning has largely been model-centric, treating model parameters as the state that changes with sequential experience. Modern agents can also adapt through a harness of prompts, memories, tools, skills, and routing rules. Because these contents jointly shape later execution, a harness update can disrupt previously reliable behavior even when the model is frozen. This raises a new question: how can an agent continually improve its state outside the model while retaining behavior acquired earlier? We formulate Harness Continual Learning (HCL), a new continual learning paradigm in which the harness evolves around a frozen foundation model, and define the resulting loss of earlier behavior as harness-level forgetting. We instantiate HCL with four execution-facing components: the Task Interface, Experience Memory, Capability Map, and Adaptive Router. We further introduce guarded harness evolution to separate update generation from state commitment. A Continual Optimizer proposes candidate harnesses from post-execution feedback, and a Continual Evaluator commits the resulting candidate harness only after checking current improvement, historical retention, and validity. Experiments on textual reasoning, multimodal perception, and open-world interaction demonstrate capability accumulation and failure recovery, with relative gains exceeding 10% over corresponding baselines in multiple settings. Component ablations assess the contribution of each harness component, while controlled retention sweeps reveal measurable harness-level forgetting and show that the stability--plasticity trade-off can be explicitly adjusted.

Borui Kang, Jinrui Gu, Junhan Lv +3

State Key Laboratory for Novel Software Technology, Nanjing University, China · University of Wollongong, Australia

Continual Learning for LLM Agents Continual Learning Agent Harness Optimization

Ask PDF

Mar 2, 2026

Position: Modular Memory is the Key to Continual Learning Agents

Foundation models have transformed machine learning through large-scale pretraining and increased test-time compute. Despite surpassing human performance in several domains, these models remain fundamentally limited in continuous operation, experience accumulation, and personalization, capabilities that are central to adaptive intelligence. While continual learning research has long targeted these goals, its historical focus on in-weight learning (IWL), i.e., updating a single model's parameters to absorb new knowledge, has rendered catastrophic forgetting a persistent challenge. Our position is that combining the strengths of In-Weight Learning (IWL) and the newly emerged capabilities of In-Context Learning (ICL) through the design of modular memory is the missing piece for continual adaptation at scale. We outline a conceptual framework for modular memory-centric architectures that leverage ICL for rapid adaptation and knowledge accumulation, and IWL for stable updates to model capabilities, charting a practical roadmap toward continually learning agents.

Vaggelis Dorovatas, Malte Schwerin, Andrew D. Bagdanov +21

Toyota Motor Europe · University of Bremen · University of Florence +17

Continual Learning for LLM Agents Memory-Augmented Agents

Ask PDF

Jun 30, 2026

AhaBench: Do Agents Turn Experience into Reusable Insights? A Long-Horizon Benchmark for Continual Learning

Can language agents continually learn from experience, turning earlier interactions into reusable capabilities? AhaBench evaluates this ability through exploration after solved hidden-state puzzles, computational transfer after mathematical teaching, and sustained business operation under delayed feedback. The benchmark is agnostic to how an agent learns; the evaluated agents use fixed model weights. Curriculum profiles, teaching contrasts, and daily trajectories reveal a common challenge: using explicit guidance is more reliable than generalizing beyond it or sustaining useful behavior. Across the Puzzle panel, the advantage over matched cold targets is 36.0-53.5 points greater with trace support than at the trace-free endpoint; Qwen 3.6 Plus nevertheless retains a +12.57-point post-curriculum gain. In Euler, worked procedures yield 80.0-100.0% held-out accuracy across models, while question-plus-answer teaching yields 0.0-73.9%. Vending trajectories separate sustained profit, late recovery, and incomplete operation: Doubao Seed 2.0 Pro finishes nominal operation at +495 but averages -10 over the year. Together, these results make continual learning an operational target: experience should yield capabilities that remain effective as guidance, inputs, and business states change. We release tasks, validators, a simulator, records, and analyses for developing agents that turn useful insights into lasting abilities.

Zerui Cheng, Jiawei Xu, Huacan Chai +3

Princeton University · Tencent Hy · Tsinghua University +2

Continual Learning for LLM Agents Lifelong Learning Agents Long-Horizon Agent Evaluation AI Agent Benchmarks Agent Skill Learning

Ask PDF

Abstract

Continual Learning for LLM Agents

Explore similar work

Harness Continual Learning: Continual Adaptation Beyond Model Parameters

Position: Modular Memory is the Key to Continual Learning Agents

AhaBench: Do Agents Turn Experience into Reusable Insights? A Long-Horizon Benchmark for Continual Learning