AuthorsLinhao Zhong, Zongze Du, Linyu Wu et al.
TL;DR
ForeDreamer uses a dual-agent MemGuide + MemTools memory architecture to turn noisy web evidence into factual memory, reaching 0.1471 Brier score on Prophet Arena vs 0.2059 for Full Text.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Open-web future event prediction requires agents to distill reliable signals from noisy, redundant, and incomplete evidence, which existing memory mechanisms cannot handle well.
Traditional agent-memory systems focus on relatively clean, user-centered context, so they struggle to process heterogeneous, conflicting web data into calibrated forecasts for tasks like Prophet Arena and FutureX.
HOW IT WORKS
ForeDreamer introduces a main agent, memory-processing subagent, factual memory, experiential memory, and an Experience Bank orchestrated via MemGuides and MemTools over an artifact workspace.
You can think of ForeDreamer like a CPU with a dedicated preprocessing coprocessor: the main agent acts as RAM for reasoning, while the subagent and MemTools behave like a disk-backed ETL pipeline organizing evidence.
This architecture lets ForeDreamer explicitly transform noisy search results into structured factual memory, enabling calibrated future predictions that a plain context window or simple RAG pipeline cannot provide.
DIAGRAM
This diagram shows how ForeDreamer runs the multi-turn search-and-reasoning loop to build factual memory and produce a forecast.
DIAGRAM
This diagram shows how ForeDreamer evolves textual forecasting experience and procedural evidence-processing experience, and how ablations disable each track.
PROCESS
Forecasting Workflow
ForeDreamer receives (q, τq), and the main agent consults the Experience Bank to plan multi-round search and reasoning while separating factual and experiential memory.
Cutoff-aware Web Search
ForeDreamer issues search queries through the search-and-process tool, retrieving Rt = Search(st, E≤τq) and writing normalized artifacts into the artifact workspace.
MemGuide and MemTools Processing
The memory-processing subagent applies a selected MemGuide and its MemTools to convert workspace artifacts into question-specific factual memory Mt via Pg,Tg(rt,k, q).
Dual-Track Experience Evolution
After forecasting, ForeDreamer logs rollout feedback to evolve the Experience Bank and expand the MemGuide tree with compositional tool reuse and diversity-guided exploration.
KEY CONTRIBUTIONS
Formalizing factual-memory management for forecasting
ForeDreamer defines factual memory as a query-conditioned evidence artifact and builds it via a memory-processing subagent using MemGuides and MemTools before prediction.
Dual-track experiential memory evolution
ForeDreamer evolves textual forecasting experience in the Experience Bank and procedural evidence-processing experience in the MemGuide tree using rollout feedback on Devo.
Compositional tool reuse and diversity-guided exploration
ForeDreamer introduces Compositional Tool Reuse and Diversity-Guided Exploration, reducing MemTool redundancy to 48.8% cluster ratio and diversifying MemGuide pipeline archetypes.
RESULTS
Prophet Arena Brier score
0.1471
-0.0588 vs Full Text on Qwen3.5-Flash
FutureX accuracy
0.4108
+0.0810 vs Full Text on Qwen3.5-Flash
GPT-5.4-Nano Prophet Brier
0.1839
-0.0245 vs Full Text with GPT-5.4-Nano
GPT-5.4-Nano FutureX accuracy
0.3883
+0.1117 vs No Information with GPT-5.4-Nano
On Prophet Arena, ForeDreamer achieves 0.1471 average Brier score with Qwen3.5-Flash, and 0.1839 with GPT-5.4-Nano, showing improved calibration over Full Text and RAG baselines. On FutureX, ForeDreamer reaches 0.4108 accuracy with Qwen3.5-Flash, demonstrating that evolving forecasting and procedural memory yields better future-event predictions than existing agent-memory systems.
BENCHMARK
On Prophet Arena, ForeDreamer achieves 0.1471 average Brier score with Qwen3.5-Flash, and 0.1839 with GPT-5.4-Nano, showing improved calibration over Full Text and RAG baselines. On FutureX, ForeDreamer reaches 0.4108 accuracy with Qwen3.5-Flash, demonstrating that evolving forecasting and procedural memory yields better future-event predictions than existing agent-memory systems.
BENCHMARK
Average Brier score on Prophet Arena; lower values indicate better calibrated forecasts.
BENCHMARK
Prediction accuracy on FutureX; higher values are better.
KEY INSIGHT
ForeDreamer with Compositional Tool Reuse collapses 201 generated MemTools into 98 clusters at token-Jaccard threshold τ = 0.5, yielding a 48.8% cluster ratio.
This is surprising because naive procedural evolution might be expected to generate highly diverse tools, but ForeDreamer reveals strong recurring patterns and redundancy in evidence-processing operations.
WHY IT MATTERS
ForeDreamer unlocks future-event agents that reason over explicit factual memory instead of raw, noisy web snippets, improving calibration and robustness in open-web forecasting.
Builders can now design agents whose memory systems self-evolve MemGuides and MemTools, turning feedback from forecasting episodes into better search strategies and evidence-processing workflows without manual tool engineering.
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…