AuthorsMinqian Liu, Shiyu Chang, Lifu Huang
TL;DR
Incremental Prompting with Episodic Memory Prompts entangles span features with type-specific prompts to reach 50.12% micro F1 on MAVEN Task 5, +10.78 over KT.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Class-incremental event detection suffers from catastrophic forgetting, where performance on old types drops sharply after training on new data.
When BERT-ED is trained to Task 5 on MAVEN, micro F1 falls to 22.69%, severely harming downstream event detection reliability over the full ontology.
HOW IT WORKS
Incremental Prompting introduces Episodic Memory Prompting, combining Span-based Event Detection, Episodic Memory Prompts, Experience Replay, and Knowledge Distillation to store and reuse type-specific knowledge.
You can think of Episodic Memory Prompts as a growing card catalog, where each new event type gets its own learnable card that stays accessible across future tasks.
By entangling span representations with all accumulated prompts, Incremental Prompting lets the classifier consult a soft episodic memory rather than relying only on a limited context window.
DIAGRAM
This diagram shows how Incremental Prompting trains on each task using Episodic Memory Prompts, experience replay, and prediction plus feature-level knowledge distillation.
DIAGRAM
This diagram shows how Incremental Prompting is evaluated on MAVEN and ACE05-EN across five class-incremental tasks with shared test sets.
PROCESS
Span-based Event Detection
Incremental Prompting freezes BERT and uses Span-based Event Detection to encode sentences and build span representations for candidate event triggers.
Episodic Memory Prompting
Incremental Prompting initializes new type-specific prompts, concatenates them with accumulated prompts, encodes them, and entangles span features with prompt representations.
Lifelong Learning with Experience Replay and Knowledge Distillation
Incremental Prompting replays exemplars from the memory buffer and applies prediction-level and feature-level Knowledge Distillation to align with the previous model.
Optimization
Incremental Prompting optimizes the combined loss L = L̃C + αLER + β(LPD + LFD) and updates the exemplar memory using the herding algorithm.
KEY CONTRIBUTIONS
Episodic Memory Prompts for class incremental event detection
Incremental Prompting introduces Episodic Memory Prompting with type-specific prompts that are carried across tasks, improving MAVEN Task 5 micro F1 from 39.34% for KT to 50.12%.
Prompt based lifelong event detection framework
Incremental Prompting is the first to apply prompt learning to class-incremental event detection, combining Span-based Event Detection, Experience Replay, and Knowledge Distillation.
Comprehensive analysis of new and old event types
Incremental Prompting analyzes per-type F1 on old and new types across tasks, showing gains on both categories compared with KT on MAVEN.
RESULTS
Micro F-1 MAVEN Task 5
50.12%
+10.78 over KT
Micro F-1 MAVEN Task 3
58.61%
+11.37 over KT
Micro F-1 ACE05 Task 5
43.28%
+0.66 over KT
Micro F-1 MAVEN Task 1
67.86%
+4.35 over BERT-ED
On MAVEN and ACE05-EN class-incremental benchmarks, Incremental Prompting is evaluated over five tasks with a shared test ontology. The gains in micro F-1, especially +10.78 on MAVEN Task 5 versus KT, show that Episodic Memory Prompts substantially reduce catastrophic forgetting while improving base event detection.
BENCHMARK
On MAVEN and ACE05-EN class-incremental benchmarks, Incremental Prompting is evaluated over five tasks with a shared test ontology. The gains in micro F-1, especially +10.78 on MAVEN Task 5 versus KT, show that Episodic Memory Prompts substantially reduce catastrophic forgetting while improving base event detection.
BENCHMARK
Micro F-1 on MAVEN Task 5 for Incremental Prompting and strong lifelong learning baselines.
BENCHMARK
Effect of prompt initialization, entangled optimization, knowledge distillation, and discrete prompts on MAVEN Task 5 micro F-1.
KEY INSIGHT
Incremental Prompting with zero exemplars (buffer size 0) still benefits from Episodic Memory Prompts, yet performance drops sharply compared with using 20 exemplars.
This is surprising because prompts are designed as a memory, but the results show that even strong prompt-based episodic memory still heavily depends on stored examples.
WHY IT MATTERS
Incremental Prompting unlocks a way to grow a soft episodic memory of event types via prompts, without expanding backbone parameters or requiring task identifiers.
Builders can now design lifelong event detectors that retain knowledge across many tasks using compact prompt vectors, making continual updates more practical under tight memory and retraining budgets.
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…