AuthorsTaeil Kim, Kangsan Kim, Sung Ju Hwang
TL;DR
Agent Memory Distillation uses hierarchical Workflow, Subtask, and Function memories to transfer GPT-5-mini experience, giving Qwen3-4B +34.52%p accuracy on AppWorld over zero-shot.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Agent Memory Distillation targets small agents whose own memories are dominated by failures, limiting gains from experience reuse despite teacher success rates.
Even when GPT-5-mini memories are naively transferred, the capability gap means students still fail tasks like multi-step tool use, capping downstream accuracy improvements.
HOW IT WORKS
Agent Memory Distillation organizes teacher experience into Workflow memory, Subtask memory, and Function memory, then retrieves them with proactive and reactive injection for the student.
You can think of Workflow memory as a high-level playbook, Subtask memory as worked examples, and Function memory as an API cookbook with concrete call patterns.
This hierarchical design lets Agent Memory Distillation give small agents structured guidance that a plain context window or flat task-level memory cannot provide.
DIAGRAM
This diagram shows how Agent Memory Distillation retrieves Workflow, Subtask, and Function memories during student inference with proactive and reactive injection.
DIAGRAM
This diagram shows how Agent Memory Distillation builds teacher memories, runs student baselines, and performs component ablations across benchmarks.
PROCESS
Hierarchical Memory Generation
Agent Memory Distillation runs the teacher on tasks to collect successful trajectories and builds Workflow memory, Subtask memory, and Function memory from these traces.
Proactive Injection of Workflow and Subtask Memories
Agent Memory Distillation retrieves a Workflow memory insight and relevant Subtask memory segments, then injects them into the student system prompt before execution.
Student Reasoning with Memory Guidance
Guided by Workflow memory and Subtask memory, the student plans tool sequences and follows concrete behavioral examples while interacting with the environment.
Reactive Function Memory Injection
When tool calls fail, Agent Memory Distillation retrieves Function memory entries for the failing function and appends targeted call examples to fix errors.
KEY CONTRIBUTIONS
Agent Memory Distillation framework
Agent Memory Distillation introduces a training free teacher to student memory transfer framework using Workflow memory, Subtask memory, and Function memory for small agents.
Hierarchical teacher memory design
Agent Memory Distillation shows that Subtask memory provides the largest gains, with Workflow memory and Function memory complementing it across task, subtask, and function levels.
Comprehensive empirical study
Agent Memory Distillation achieves up to 54.17% accuracy on AppWorld and average gains of 27.2%p, 11.2%p, and 3.4%p on AppWorld, BFCL V3, and ToolSandbox.
RESULTS
AppWorld accuracy
49.40%
+34.52%p over Qwen3-4B Zero-shot
BFCL V3 accuracy
38.50%
+23.00%p over Qwen3-4B Zero-shot
ToolSandbox accuracy
20.16%
+3.88%p over Qwen3-4B Zero-shot
Average accuracy
36.02%
+20.47%p over Qwen3-4B Zero-shot
On AppWorld, BFCL V3, and ToolSandbox, which test multi app Python APIs, structured function calling, and stateful tool use, Agent Memory Distillation consistently raises small agent accuracy far above zero shot baselines.
BENCHMARK
On AppWorld, BFCL V3, and ToolSandbox, which test multi app Python APIs, structured function calling, and stateful tool use, Agent Memory Distillation consistently raises small agent accuracy far above zero shot baselines.
BENCHMARK
Accuracy (%) on AppWorld test_normal comparing Agent Memory Distillation to memory based baselines using teacher trajectories.
BENCHMARK
Accuracy (%) on AppWorld as Agent Memory Distillation adds Workflow, Subtask, and Function memories.
KEY INSIGHT
For Qwen3-4B on AppWorld, GPT-5-mini as teacher yields 49.40% accuracy, beating DeepSeek V4 Pro despite DeepSeek’s higher 81.55% teacher accuracy.
This is surprising because one might expect the strongest teacher, GPT-5.5 at 91.08% accuracy, to always transfer best, but teacher student compatibility matters more than raw teacher strength.
WHY IT MATTERS
Agent Memory Distillation shows that small 4B to 8B agents can reach or surpass GPT-5-mini performance using only hierarchical teacher memories.
Builders can now deploy cheaper small agents that still follow rich multi level tool use strategies without any student fine tuning or parameter updates.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
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.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…