AuthorsZitong Shi, Yixuan Tang, Anthony Kum Hoe Tung
TL;DR
A-TMA adds bank, retrieval, and QA state-role alignment to long-term agent memory, raising Graphiti/Zep conflict accuracy on LTP from 0.480 to 0.720 (+0.240).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Ghost memory occurs when old, current, and transition facts coexist in the bank and mislead QA, even though each fact was once true.
Graphiti/Zep reaches only 0.480 conflict accuracy on LTP, so long term assistants can ship packages to outdated addresses or follow obsolete preferences.
HOW IT WORKS
A-TMA combines Bank Level State Maintenance, Retrieve Level Evidence Construction, and QA Level Evidence State Conditioning around host systems like Graphiti/Zep.
You can think of A-TMA as adding a versioned index and labels on top of RAM, so queries see the right snapshot instead of a mixed log.
This explicit state-role overlay lets A-TMA answer from the requested state view, something a plain context window or timestamp-only memory cannot guarantee.
DIAGRAM
This diagram shows how A-TMA routes a single query through query state view detection, state aligned retrieval, and labeled QA resolution.
DIAGRAM
This diagram shows how A-TMA is evaluated on LTP and LoCoMo, from profiles and conversations to bank, retrieval, and QA metrics.
PROCESS
Bank Level State Maintenance
A-TMA uses Bank Level State Maintenance to keep active, superseded, and transition records, linking them with supersedes and evolves relations instead of deleting history.
Sentry candidate proposal
During Sentry candidate proposal, A-TMA scores new memories against existing ones to find suspicious same slot pairs that may indicate state changes.
Retrieve Level Evidence Construction
In Retrieve Level Evidence Construction, A-TMA expands host retrieval with relation hops and builds a candidate pool aligned with the query state view.
QA Level Evidence State Conditioning
QA Level Evidence State Conditioning serializes evidence with labels like current memory and historical memory, then constrains QA to answer from the requested state.
KEY CONTRIBUTIONS
Three level state changing memory formulation
A-TMA formalizes state changing memory as Bank Level State Maintenance, Retrieve Level Evidence Construction, and QA Level Evidence State Conditioning, decoupling failures hidden by final QA accuracy.
Ghost memory as state coordination failure
A-TMA identifies ghost memory where old, current, and transition facts remain mixed, and uses Bank Level State Maintenance plus Sentry candidate proposal to preserve history without stale answers.
LTP benchmark and host dependent gains
A-TMA introduces LTP with 10 profiles and 800 probes, where Graphiti/Zep conflict accuracy rises from 0.480 to 0.720 and temporal F1 on LoCoMo increases from 0.0295 to 0.1705.
RESULTS
Conflict accuracy
0.720
+0.240 over Graphiti/Zep on LTP
QA Acc.
0.635
+0.111 over Graphiti/Zep on LTP
Temporal F1
0.1705
+0.141 over Graphiti/Zep on LoCoMo temporal
Average F1
0.1556
+0.0747 over Graphiti/Zep on LoCoMo average
On LTP, which stresses conflicting old and current facts, A-TMA lifts Graphiti/Zep conflict accuracy from 0.480 to 0.720 and QA accuracy from 0.524 to 0.635. On LoCoMo, which tests long conversation generalization, A-TMA raises Graphiti/Zep temporal F1 from 0.0295 to 0.1705 and average F1 from 0.0809 to 0.1556.
BENCHMARK
On LTP, which stresses conflicting old and current facts, A-TMA lifts Graphiti/Zep conflict accuracy from 0.480 to 0.720 and QA accuracy from 0.524 to 0.635. On LoCoMo, which tests long conversation generalization, A-TMA raises Graphiti/Zep temporal F1 from 0.0295 to 0.1705 and average F1 from 0.0809 to 0.1556.
BENCHMARK
Conflict accuracy on LTP, measuring state correct answers when old and current facts conflict.
KEY INSIGHT
Graphiti/Zep reaches 0.812 evidence support on LTP but only 0.425 conflict accuracy, while A-TMA raises conflict accuracy to 0.725.
This is surprising because timestamped temporal graphs seem state aware, yet A-TMA shows that explicit state roles are still required for answer time correctness.
WHY IT MATTERS
A-TMA lets builders keep full state history while still answering from the correct current or historical view for each query.
Developers can now plug A-TMA into existing systems like Graphiti/Zep or InsideOut to debug bank, retrieval, and QA failures separately instead of guessing from a single QA score.
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…