AuthorsYuxi Qian, Yuxiang Ren
TL;DR
EvoGraph-Mem adds failure-aware editable insight nodes with positive and negative evidence, yielding 47.47 avg score vs 43.23 for G-Memory (+4.24).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
EvoGraph-Mem targets the issue that append-only memory is insufficient for long-horizon tasks, where reused insights become outdated, conflicting, or harmful.
When polluted memories are repeatedly retrieved, long-term language agents face degraded reasoning quality and task performance across evolving tasks and environments.
HOW IT WORKS
EvoGraph-Mem combines a Graph-Based Historical Memory, an Insight Graph, Utility-Aware Insight Retrieval, and a Graph Controller for Memory Correction to maintain insight-level validity.
You can think of EvoGraph-Mem like a library card catalog that not only files books but also flags, revises, or archives misleading entries over time.
This design lets EvoGraph-Mem selectively reuse only well-supported insights, something a plain context window or append-only memory cannot safely guarantee.
DIAGRAM
This diagram shows how EvoGraph-Mem retrieves insights with evidence-aware scoring and then edits the insight graph after task execution.
DIAGRAM
This diagram shows how EvoGraph-Mem is evaluated on three benchmarks and compared against memory-based baselines.
PROCESS
Graph-Based Historical Memory
EvoGraph-Mem first builds the Graph-Based Historical Memory with a structured Query Graph and Insight Graph over past tasks and interactions.
Utility-Aware Insight Retrieval
Given a new query, EvoGraph-Mem performs Utility-Aware Insight Retrieval, scoring candidates by positive evidence st,k and conflict penalty λ ct,k.
Failure-Aware Insight Representation
During execution, EvoGraph-Mem uses Failure-Aware Insight Representation with Ω+ k, Ω− k, and activation state zk to track how each insight behaves.
Graph Controller for Memory Correction
After task feedback, the Graph Controller for Memory Correction issues KEEP, ARCHIVE, REVISE, and ADD operations to evolve the Insight Graph.
KEY CONTRIBUTIONS
Insight-level memory maintenance
EvoGraph-Mem formalizes insight-level memory maintenance, showing append-only memory is insufficient and linking polluted insights to degraded long-horizon performance on PDDL, HotpotQA, and FEVER.
Editable insight graph
EvoGraph-Mem introduces an editable Insight Graph where each node stores positive evidence Ω+ k, negative evidence Ω− k, and activation state zk for failure-aware retrieval.
Graph controller for memory correction
EvoGraph-Mem adds a Graph Controller for Memory Correction that issues KEEP, ARCHIVE, REVISE, and ADD operations, yielding a 21.75 exact match gain on HotpotQA over the no-memory baseline with GPT-4o-mini.
RESULTS
PDDL progress rate GPT-4o-mini
30.67
+2.90 over G-Memory
HotpotQA exact match GPT-4o-mini
43.43
+7.76 over G-Memory
FEVER exact match GPT-4o-mini
68.32
+2.08 over G-Memory
Average score GPT-4o-mini
47.47
+4.24 over G-Memory
On PDDL, HotpotQA, and FEVER, EvoGraph-Mem is evaluated using progress rate and exact match accuracy, directly testing long-horizon planning and multi-hop reasoning. The 47.47 vs 43.23 average score with GPT-4o-mini shows EvoGraph-Mem converts corrective memory maintenance into consistent benchmark gains over G-Memory.
BENCHMARK
On PDDL, HotpotQA, and FEVER, EvoGraph-Mem is evaluated using progress rate and exact match accuracy, directly testing long-horizon planning and multi-hop reasoning. The 47.47 vs 43.23 average score with GPT-4o-mini shows EvoGraph-Mem converts corrective memory maintenance into consistent benchmark gains over G-Memory.
BENCHMARK
Average score across PDDL, HotpotQA, and FEVER with GPT-4o-mini.
BENCHMARK
Average token consumption for GPT-4o-mini across all benchmarks.
KEY INSIGHT
EvoGraph-Mem increases average token usage to 6.4M on GPT-4o-mini, a 73.0% rise over the 3.7M no-memory baseline.
Despite this overhead, EvoGraph-Mem still improves the average score from 36.41 to 47.47, challenging the assumption that tighter token budgets always yield better efficiency–performance trade-offs.
WHY IT MATTERS
EvoGraph-Mem unlocks long-term agents that can not only remember but also correct, archive, and refine their own high-level insights over time.
Builders can now deploy language agents whose external memory evolves through explicit KEEP, ARCHIVE, REVISE, and ADD operations, avoiding memory pollution in complex, evolving 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…