AuthorsShuqi Cao, Jingyi He, Fei Tan
TL;DR
HiGMem uses an LLM-guided event–turn hierarchy to filter memories, boosting adversarial F1 from 0.54 to 0.78 on LoCoMo10 while retrieving 8.09 vs 99.84 turns.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory systems relying on vector similarity produce bloated evidence sets, where extra similar turns yield little recall but erode precision and inflate context.
In long-term conversational agents, this makes retrieval precision drop, raises answer-stage context cost, and leaves evidence harder to inspect and manage over extended histories.
HOW IT WORKS
HiGMem’s core mechanism combines a Turn Layer, Event Layer, Memory Construction Mechanism, and LLM-Guided Retrieval to structure and filter conversational memories.
You can think of HiGMem like a library card catalog: events are shelves with concise cards, and turns are the individual books the librarian pulls on demand.
This event–turn hierarchy lets HiGMem reason over summaries to pick only turns worth reading, something a plain context window or pure vector store cannot achieve.
DIAGRAM
This diagram shows how HiGMem constructs Turn nodes, affiliates them to Event nodes, and updates summaries and links when a new dialogue turn arrives.
DIAGRAM
This diagram shows how HiGMem processes a query, retrieves semantic events and turns, predicts useful turns, filters candidates, and generates the final response.
PROCESS
Memory Construction Mechanism
HiGMem uses the Memory Construction Mechanism to analyze each new dialogue turn with a sliding window and build enriched Turn nodes with metadata.
Event Affiliation and Update
HiGMem performs Event Affiliation and Update by embedding Turn nodes, selecting top-kevent Event candidates, and refreshing or appending summaries and fact sheets based on volume threshold τ.
LLM-Guided Retrieval
During LLM-Guided Retrieval, HiGMem generates query keywords, retrieves semantic Turn and Event nodes, and lets the LLM reason over event-linked turns to form Tpred.
Final Evidence Set and Response
HiGMem merges Tsemantic and Tpred into Tcand, applies an LLM-based Filter to obtain Tfinal, and then feeds Tfinal with the query into GPT-4o-mini or GPT-5 to generate the final response.
KEY CONTRIBUTIONS
Hierarchical event–turn memory architecture
HiGMem introduces a two-level Turn Layer and Event Layer with explicit links, enabling multi-granularity reasoning and improving F1 from 0.39 to 0.49 over a non-hierarchical variant.
LLM-guided Hierarchical Retrieval strategy
HiGMem’s LLM-Guided Retrieval uses event summaries as semantic anchors to predict useful turns, reaching adversarial F1 0.78 vs 0.54 for A-Mem on LoCoMo10.
Efficient long-term conversational memory baseline
HiGMem provides an open-source system that retrieves on average 8.09 turns with Precision@K 0.1909 while keeping Recall@K 0.7241, compared to A-Mem’s 99.84 turns and Precision@K 0.0101.
RESULTS
Adversarial F1
0.78
+0.24 over A-Mem
Single-Hop F1
0.49
+0.07 over A-Mem
Avg K
8.09 turns
−91.75 turns vs A-Mem
Precision@K
0.1909
+0.1808 over A-Mem
On the LoCoMo10 benchmark, which tests long multi-session conversational QA across Single-Hop, Multi-Hop, Temporal, Open-Domain, and Adversarial questions, HiGMem shows that hierarchical LLM-guided memory can raise adversarial F1 to 0.78 while shrinking the evidence set from 99.84 to 8.09 turns.
BENCHMARK
On the LoCoMo10 benchmark, which tests long multi-session conversational QA across Single-Hop, Multi-Hop, Temporal, Open-Domain, and Adversarial questions, HiGMem shows that hierarchical LLM-guided memory can raise adversarial F1 to 0.78 while shrinking the evidence set from 99.84 to 8.09 turns.
BENCHMARK
F1 score on LoCoMo10 adversarial questions.
BENCHMARK
Average number of turns in the final evidence set (Avg K).
KEY INSIGHT
HiGMem retrieves only 8.09 turns on average yet maintains Recall@K 0.7241, nearly matching A-Mem’s 0.7502 with 99.84 turns.
This is surprising because many assume higher recall requires flooding the answer model with more context, but HiGMem shows LLM-guided selection can keep recall high with far fewer turns.
WHY IT MATTERS
HiGMem unlocks compact, high-precision evidence retrieval for long-term conversational agents, enabling multi-session reasoning without overwhelming answer-stage context windows.
Builders can now design agents that maintain rich long-term memory while keeping token cost and latency manageable, making hybrid deployments with expensive answer models like GPT-5 practical.
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…