AuthorsYiheng Shu, Bernal Jiménez Gutiérrez, Saisri Padmaja Jonnalagedda et al.
TL;DR
AGENTCL uses controlled compositional task streams plus MEMPROBE’s multi-view memory probing to expose large plasticity gains (e.g., +21.9 pp on CodeEval-Pro) that naive streams hide.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Language agents spend substantial inference time on tasks, yet experience acquired in one episode is often underutilized in future episodes.
Existing benchmarks use naive task streams and average accuracy, making it hard to see plasticity, stability, and what is actually being reused over time.
HOW IT WORKS
AGENTCL introduces naive streams, compositional streams, and a two pass protocol, while MEMPROBE builds interaction memory, insight memory, and skill memory with validation.
Think of AGENTCL as a carefully scripted curriculum and MEMPROBE as a librarian that files episodes as raw logs, distilled lessons, and reusable code snippets.
This design lets AGENTCL measure plasticity, stability, and generalization in ways a plain context window or uncontrolled task order cannot.
DIAGRAM
This diagram shows how AGENTCL runs first pass and second pass over the same task stream to compute Plasticity Gain and Stability Gain.
DIAGRAM
This diagram shows how AGENTCL builds naive and compositional streams, runs two pass evaluation, and then tests held out tasks.
PROCESS
Task Streams
AGENTCL constructs naive and compositional task streams where later tasks may reuse earlier sub solutions, evidence, or workflows.
Two Passes
AGENTCL runs a first pass with read write memory and a second pass with frozen memory to obtain Fi and Si for each task.
Transfer Metrics
AGENTCL computes Plasticity Gain PGi = Fi − Bi, Stability Gain SGi = Si − Fi, and Generalization Gain GGj = Hj − Bj.
MEMPROBE Loop
MEMPROBE performs retrieve solve consolidate, building interaction memory, insight memory, and skill memory while filtering unreliable experiences.
KEY CONTRIBUTIONS
AGENTCL evaluation framework
AGENTCL formalizes continual learning in agents with naive versus compositional task streams and explicit Plasticity Gain, Stability Gain, and Generalization Gain metrics.
Controlled task streams
AGENTCL constructs compositional streams on CodeEval-Pro and BrowseComp+ where 48 and 308 subtasks respectively precede 48 and 100 complex tasks.
MEMPROBE memory probing
AGENTCL introduces MEMPROBE with interaction memory, insight memory, and skill memory plus validation, achieving large compositional PG like +21.9 pp on CodeEval-Pro.
RESULTS
Average accuracy MMLU Pro 1st Pass
82.7%
+4.0 over ReAct
Average accuracy MMLU Pro 2nd Pass
83.0%
+4.3 over ReAct
Std Dev accuracy MMLU Pro 1st Pass
1.7
method separation across ReAct, ExpRAG, ReMem, MEMPROBE
Std Dev accuracy MMLU Pro 2nd Pass
1.8
shows naive streams compress differences
These numbers come from the naive MMLU-Pro streams in AGENTCL, where each domain has 100 tasks and memory is maintained separately. The results show MEMPROBE slightly improves over ReAct, while small standard deviations highlight that naive streams provide limited discriminative power for continual learning.
BENCHMARK
These numbers come from the naive MMLU-Pro streams in AGENTCL, where each domain has 100 tasks and memory is maintained separately. The results show MEMPROBE slightly improves over ReAct, while small standard deviations highlight that naive streams provide limited discriminative power for continual learning.
BENCHMARK
Average accuracy across economics, engineering, and philosophy on the naive MMLU-Pro stream first pass.
KEY INSIGHT
On BrowseComp+, compositional streams yield accuracy standard deviations of 14.9 and 16.0, while naive streams show only 2.3 and 5.7 across methods.
This is surprising because many assume streaming tasks from the same domain is enough, but AGENTCL shows explicit compositional structure is crucial for revealing plasticity differences.
WHY IT MATTERS
AGENTCL lets researchers precisely measure how language agents accumulate and reuse experience using MEMPROBE’s validated interaction, insight, and skill memories.
Builders can now design and debug continual learning agents with clear plasticity, stability, and generalization signals instead of relying on opaque average scores from uncontrolled task streams.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…