AuthorsJunfeng Liao, Qizhou Wang, Jianing Zhu et al.
TL;DR
BeliefMem uses probabilistic candidate memories with noisy OR updates to achieve 42.38 F1 on LoCoMo vs 40.99 for Mem0 (+1.39).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing agents store each observation as a single deterministic conclusion, permanently discarding uncertainty and alternative hypotheses about hidden state.
When memory records only “API X failed,” the agent avoids testing API X, reinforcing incorrect beliefs and conflicting with user instructions like “Use API X to ...”.
HOW IT WORKS
BeliefMem introduces an external Belief Memory Bank, with Add, Merge, and Belief-aware Retrieval to maintain attribute-level candidate hypotheses and probabilities.
Think of BeliefMem like a card catalog of beliefs: instead of one card per fact, it keeps multiple cards per attribute, each with a confidence score that updates over time.
This probabilistic representation lets BeliefMem preserve uncertainty beyond a plain context window, enabling agents to revisit alternatives and correct earlier errors as new evidence arrives.
DIAGRAM
This diagram shows how BeliefMem updates candidate probabilities with Add and Merge, then retrieves top K attributes with time decay for action selection.
DIAGRAM
This diagram shows how BeliefMem is plugged into LLM agents and evaluated on LoCoMo and ALFWorld with baselines.
PROCESS
Belief-based Memory Formulation
BeliefMem first represents each attribute c with a belief over hypotheses H(c), approximating b_t(c) using the Belief Memory Bank instead of point estimates.
Belief Update in Memory
BeliefMem applies Add to initialize candidates for new attributes and Merge with noisy OR to update p_t+1(h) based on evidence strength Δ(o_t+1, h).
Belief-aware Retrieval
BeliefMem scores attributes with α_t(c) = sim(o_t, c) · λ^{τ_t(c)}, selects top K, and returns full candidate distributions p_t(h) for decision making.
Action Selection under Partial Observability
The agent policy π chooses action a_t conditioned on current observation o_t and retrieved beliefs r_t, keeping alternative hypotheses visible to avoid self-reinforcing error.
KEY CONTRIBUTIONS
Belief-based Memory Representation
BeliefMem replaces deterministic entries with attribute-level belief states in the Belief Memory Bank, storing multiple candidate conclusions per attribute with probabilities p_t(h).
Noisy OR Evidence Merge
BeliefMem introduces a noisy OR Merge update p_t+1(h) = min(1 − (1 − p_t(h))(1 − Δ(o_t+1, h)), 0.99), enabling robust belief refinement under noisy observations.
Empirical Gains on LoCoMo and ALFWorld
BeliefMem achieves 42.38 F1 on LoCoMo with GPT-4o-mini and 59.88% SR on ALFWorld, exceeding Mem0 by 1.39 F1 and ReadAgent by 5.85 SR.
RESULTS
LoCoMo F1 Avg GPT4o mini
42.38 F1
+1.39 over Mem0 (40.99 F1)
LoCoMo BLEU1 Avg GPT4o mini
36.30 BLEU-1
+4.31 over Mem0 (31.99 BLEU-1)
ALFWorld Avg SR
59.88%
+5.85 over ReadAgent (54.03% SR)
ALFWorld Avg Steps
29.56 steps
- -3.99 vs ReadAgent (27.65 steps, more steps for higher SR)
BeliefMem is evaluated on LoCoMo for long-term conversational memory and ALFWorld for embodied interaction. These results show BeliefMem’s probabilistic memory improves multi-hop reasoning and success rate under partial observability compared to deterministic baselines.
BENCHMARK
BeliefMem is evaluated on LoCoMo for long-term conversational memory and ALFWorld for embodied interaction. These results show BeliefMem’s probabilistic memory improves multi-hop reasoning and success rate under partial observability compared to deterministic baselines.
BENCHMARK
Average F1 on LoCoMo across all question categories using GPT-4o-mini.
BENCHMARK
Average success rate on ALFWorld seen and unseen splits with Qwen3-Next-80B-A3B-Instruct.
KEY INSIGHT
BeliefMem using only 50% of the ALFWorld memory corpus (BeliefMem*) still achieves 59.88% average SR, beating all baselines including ReadAgent at 54.03%.
This is counterintuitive because more memory data usually helps, yet BeliefMem’s probabilistic representation avoids overfitting seen trajectories and improves out-of-distribution generalization with less data.
WHY IT MATTERS
BeliefMem unlocks agents that can act under partial observability while keeping multiple hypotheses alive, updating beliefs instead of locking into brittle deterministic conclusions.
Builders can now design long-lived LLM agents whose memory corrects itself over time, supports temporal reasoning, and resists self-reinforcing errors in noisy, real-world environments.
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…