ForeDreamer: A Self-Evolving Dual-Agent Memory Architecture for Future Event Prediction

Linhao Zhong; Zongze Du; Linyu Wu; Yu Bo; Hourong Li; Chenchen Jing; Hao Chen; Yuling Xi

cs.CLAug 21, 2026

ForeDreamer: A Self-Evolving Dual-Agent Memory Architecture for Future Event Prediction

Authors: Linhao Zhong, Zongze Du, Linyu Wu, Yu Bo, Hourong Li, Chenchen Jing, Hao Chen, Yuling Xi,

Organizations: Zhejiang University, State Lab of CAD & CG · Ant Group · National University of Singapore · Zhejiang University of Technology

Abstract

Open-web future event prediction requires agents to distill reliable signals from noisy, redundant, and incomplete evidence. Existing retrieval/memory mechanisms directly feed retrieved information to agents or rely on simple memory functions such as storing and reusing prior information for prediction, leaving them insufficient for open-web forecasting. We propose to transform raw web evidence into structured memory before prediction, enabling agents to reason over distilled, question-specific evidence rather than noisy retrieval results. This paper presents ForeDreamer, a self-evolving dual-agent framework for managing memory over open-web evidence. ForeDreamer separates factual memory, a question-specific evidence state for the current forecast, from experiential memory, persistent agent experience accumulated across forecasting episodes. It uses a main agent for search and prediction, and a memory-processing subagent to convert search results into factual memory with dedicated tools. ForeDreamer further evolves experiential memory through two tracks, improving both forecasting decisions and factual-memory construction. Experiments on Prophet Arena and FutureX demonstrate the effectiveness of ForeDreamer. Project page: https://zhongzero.github.io/ForeDreamer

Figures & tables

Figure 1: Task mismatch of existing agent-memory systems. Traditional memory tasks often involve relatively clean, user-centered context and can be addressed by storing, retrieving, or reusing prior information. Open-web future event prediction instead requires agents to handle heterogeneous, noisy, and conflicting web data, which existing memory mechanisms are insufficient to process into forecast-ready evidence.

Figure 2: Overview of ForeDreamer. (a) During forecasting, the main agent uses forecasting experience from the Experience Bank, invokes a search-and-process tool to collect temporally valid evidence, and receives processed factual memory from a memory-processing subagent. The subagent follows a MemGuide and uses executable MemTools through an artifact workspace. (b) Rollout feedback drives dual-track self-evolution: textual forecasting experience evolves through validated edits to the Experience Bank, while procedural evidence-processing experience evolves through validated updates to the MemGuide tree and MemTools.

Figure 3: Evolution of procedural evidence-processing experience. (a) Vanilla rollout-based expansion grows a MemGuide tree by selecting a parent MemGuide, sampling an evolution rollout, deriving a design requirement, generating a candidate MemGuide and MemTools, and admitting the candidate through a validation gate. (b) Compositional tool reuse optimizes the candidate-generation step by selecting compatible existing MemTools before generating the MemGuide and any missing MemTools. (c) Diversity-guided exploration uses an evolution scheduler to retain the rollout-based path while adding an exploration path that generates category-diverse candidates from MemGuide-category summaries and representative MemGuides without requiring a new rollout result.

Figure 4: Hierarchical clustering of MemTools generated by vanilla procedural evolution on FutureX, using token-Jaccard source-code similarity. The reduced cluster count indicates that many generated MemTools share substantial token-level overlap, suggesting that vanilla evolution spends many attempts on near-duplicate operations.

Figure 5: Pipeline archetype distribution for vanilla rollout-based expansion on FutureX. All 10 evolved MemGuides fall into the same tool-orchestrated processing archetype, indicating limited exploration of alternative evidence-processing pipelines.

Model	Method	Climate/Weather	Companies	Economics	Entertainment	Mentions	Other	Politics	Sports	Avg. ( $\downarrow$ )
Qwen3.5-Flash	Full Text	0.3342	0.2657	0.0720	0.2005	0.2106	0.1306	0.2214	0.2122	0.2059
	RAG	0.2511	0.2319	0.0752	0.1739	0.2178	0.1381	0.2094	0.2030	0.1876
	HippoRAG 2	0.2798	0.2553	0.0524	0.1318	0.2020	0.1423	0.2069	0.1945	0.1831
	Mem0	0.2792	0.2902	0.0695	0.1936	0.2308	0.1778	0.2423	0.2045	0.2110
	MemoryOS	0.2385	0.2348	0.0542	0.1155	0.1917	0.1676	0.2216	0.1925	0.1771
	A-MEM	0.2306	0.2591	0.0602	0.1516	0.2218	0.1542	0.2122	0.1996	0.1862

Table 1: Main comparison on Prophet Arena. The table reports Brier score for each category and the average score; lower values indicate better calibrated forecasts.

Model	Method	Accuracy ( $\uparrow$ )
Qwen3.5-Flash	Full Text	0.3298
	RAG	0.3382
	HippoRAG 2	0.3285
	Mem0	0.3264
	MemoryOS	0.3269
	A-MEM	0.3495

Table 2: Main comparison on FutureX. The table reports prediction accuracy; higher values are better.

Method	Prophet Arena Avg. ( $\downarrow$ )	FutureX ( $\uparrow$ )
Full Text	0.2059	0.3298
w/o evolving MemGuide&MemTool	0.1663	0.3892
w/o evolving Experience Bank	0.1769	0.3351
ForeDreamer	0.1471	0.4108

Table 3: Dual-track experience ablation on Qwen3.5-Flash. Prophet Arena reports average Brier score, while FutureX reports accuracy.

Method	Prophet Arena Avg. ( $\downarrow$ )	FutureX ( $\uparrow$ )
Full Text	0.2059	0.3298
w/o Both Optimizations	0.1592	0.3850
w/o Compositional Tool Reuse	0.1541	0.4032
w/o Diversity-Guided Exploration	0.1554	0.3564
ForeDreamer	0.1471	0.4108

Table 4: Procedural-evolution optimization ablation on Qwen3.5-Flash. The variants remove Compositional Tool Reuse, Diversity-Guided Exploration, or both; Prophet Arena reports average Brier score, and FutureX reports accuracy.

Setting	Prophet Arena Full Text ( $\downarrow$ )	Prophet Arena ForeDreamer ( $\downarrow$ )	FutureX Full Text ( $\uparrow$ )	FutureX ForeDreamer ( $\uparrow$ )
Base Setting	0.2059	0.1471	0.3298	0.4108
Max Interaction Turns = 3	0.1871	0.1570	0.3894	0.4423
Max Interaction Turns = 4	0.1908	0.1629	0.3942	0.4279
Firecrawl Search	0.1995	0.1514	0.4394	0.4596
Top-2 Retrieved Results	0.1831	0.1514	0.3544	0.3606
Top-6 Retrieved Results	0.1845	0.1426	0.3398	0.4183

Table 5: Robustness across search settings on Qwen3.5-Flash. Each row changes one search or interaction configuration from the base setup and reports the corresponding comparison; Prophet Arena reports average Brier score, and FutureX reports accuracy.

Prophet Arena Avg. ( $\downarrow$ )
Model	No Information	Full Text	ForeDreamer
Qwen3.5-Flash	0.2545	0.2059	0.1471
GPT-5.4-Nano	0.2492	0.2084	0.1839
FutureX ( $\uparrow$ )
Qwen3.5-Flash	0.3077	0.3298	0.4108
GPT-5.4-Nano	0.2260	0.2766	0.3883

Table 6: Check for potential data leakage through no-information prompting. The table compares forecasting without web search against settings that use retrieved evidence.

Figure 6: MemTool similarity analysis on FutureX using token-Jaccard source-code similarity. Compositional Tool Reuse produces a less redundant tool set under the same clustering criterion.

Figure 7: MemGuide pipeline-archetype analysis on FutureX. Diversity-Guided Exploration encourages evolved guides to cover more pipeline archetypes, while vanilla rollout-based expansion remains concentrated in one archetype.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Method	Climate/Weather	Companies	Economics	Entertainment	Mentions	Other	Politics	Sports	Avg. ( $\downarrow$ )
GPT-5-Nano	Full Text	0.1994	0.2870	0.1347	0.1989	0.2225	0.1897	0.2537	0.2066	0.2116
	RAG	0.2438	0.2489	0.0988	0.1884	0.1997	0.1759	0.2658	0.2049	0.2033
	HippoRAG 2	0.2250	0.2617	0.0987	0.2089	0.2437	0.1656	0.2607	0.2024	0.2083
	Mem0	0.2124	0.3497	0.1373	0.2242	0.2024	0.2378	0.2751	0.2025	0.2302
	MemoryOS	0.1366	0.2603	0.1354	0.2021	0.2412	0.2036	0.2598	0.2063	0.2057
	A-MEM	0.2189	0.2786	0.1075	0.1935	0.2142	0.2021	0.2688	0.2032	0.2109

Appendix

Table S1: Additional Prophet Arena results for GPT-5-Nano. The table reports Brier score for each category and the average score; lower values indicate better calibrated forecasts.

Model	Method	Accuracy ( $\uparrow$ )
GPT-5-Nano	Full Text	0.3723
	RAG	0.3942
	HippoRAG 2	0.3894
	Mem0	0.3776
	MemoryOS	0.3667
	A-MEM	0.3510

Appendix

Table S2: Additional FutureX results for GPT-5-Nano. The table reports prediction accuracy; higher values are better.

Method	Market Return ( $\uparrow$ )
Full Text	0.5790
RAG	0.5798
HippoRAG 2	0.6038
Mem0	0.4763
MemoryOS	0.4901
A-MEM	0.5679

Appendix

Table S3: Additional Prophet Arena average market-return results across categories on Qwen3.5-Flash; higher values are better.

Method	Climate/Weather	Companies	Economics	Entertainment	Mentions	Other	Politics	Sports	Avg. ( $\downarrow$ )
Full Text	0.3342	0.2657	0.0720	0.2005	0.2106	0.1306	0.2214	0.2122	0.2059
w/o evolving MemGuide&MemTool	0.2328	0.2199	0.0800	0.1125	0.1678	0.1274	0.1843	0.2056	0.1663
w/o evolving Experience Bank	0.2440	0.2492	0.0666	0.1351	0.1869	0.1306	0.2035	0.1996	0.1769
ForeDreamer	0.1770	0.2162	0.0414	0.0962	0.1573	0.1274	0.1742	0.1869	0.1471

Appendix

Table S4: Additional dual-track experience ablation on Prophet Arena for Qwen3.5-Flash. The table reports Brier score; lower values are better.

Method	Climate/Weather	Companies	Economics	Entertainment	Mentions	Other	Politics	Sports	Avg. ( $\downarrow$ )
Full Text	0.2275	0.3719	0.0631	0.1424	0.2878	0.1410	0.2257	0.2075	0.2084
w/o evolving MemGuide&MemTool	0.2275	0.3719	0.0963	0.1349	0.2878	0.1397	0.2222	0.2003	0.2101
w/o evolving Experience Bank	0.2251	0.3456	0.0547	0.1551	0.2915	0.1546	0.2230	0.2075	0.2071
ForeDreamer	0.1811	0.3456	0.0506	0.1117	0.2704	0.1272	0.1927	0.1919	0.1839

Appendix

Table S5: Additional dual-track experience ablation on Prophet Arena for GPT-5.4-Nano. The table reports Brier score; lower values are better.

Method	Accuracy ( $\uparrow$ )
Full Text	0.2766
w/o evolving MemGuide&MemTool	0.2766
w/o evolving Experience Bank	0.3404
ForeDreamer	0.3883

Appendix

Table S6: Additional dual-track experience ablation on FutureX for GPT-5.4-Nano. The table reports prediction accuracy; higher values are better.

Method	Climate/Weather	Companies	Economics	Entertainment	Mentions	Other	Politics	Sports	Avg. ( $\downarrow$ )
Full Text	0.3342	0.2657	0.0720	0.2005	0.2106	0.1306	0.2214	0.2122	0.2059
w/o Both Optimizations	0.2319	0.1948	0.0471	0.1096	0.1917	0.1169	0.1881	0.1935	0.1592
w/o Compositional Tool Reuse	0.1531	0.2149	0.0621	0.1038	0.1890	0.1226	0.1965	0.1907	0.1541
w/o Diversity-Guided Exploration	0.1747	0.1935	0.0602	0.1023	0.1823	0.1439	0.1921	0.1942	0.1554
ForeDreamer	0.1770	0.2162	0.0414	0.0962	0.1573	0.1274	0.1742	0.1869	0.1471

Appendix

Table S7: Additional procedural-evolution optimization ablation on Prophet Arena for Qwen3.5-Flash. The table reports Brier score; lower values are better.

Setting	Method	Climate/Weather	Companies	Economics	Entertainment	Mentions	Other	Politics	Sports	Avg. ( $\downarrow$ )
Base Setting	Full Text	0.3342	0.2657	0.0720	0.2005	0.2106	0.1306	0.2214	0.2122	0.2059
Base Setting	ForeDreamer	0.1770	0.2162	0.0414	0.0962	0.1573	0.1274	0.1742	0.1869	0.1471
Max Interaction Turns = 3	Full Text	0.2577	0.2437	0.0955	0.1354	0.2051	0.1639	0.2016	0.1941	0.1871
Max Interaction Turns = 3	ForeDreamer	0.2165	0.2227	0.0655	0.0930	0.1741	0.1286	0.1710	0.1844	0.1570
Max Interaction Turns = 4	Full Text	0.2838	0.2196	0.1283	0.1830	0.1999	0.1408	0.1814	0.1897	0.1908
Max Interaction Turns = 4	ForeDreamer	0.2480	0.2276	0.0538	0.0992	0.1672	0.1263	0.1974	0.1834	0.1629

Appendix

Table S8: Per-category results for additional search settings on Prophet Arena with Qwen3.5-Flash (Brier score; lower is better).

Method	Climate/Weather	Companies	Economics	Entertainment	Mentions	Other	Politics	Sports	Avg. ( $\downarrow$ )
No Information	0.2576	0.3837	0.1225	0.2634	0.2576	0.2390	0.2945	0.2177	0.2545
Full Text	0.3342	0.2657	0.0720	0.2005	0.2106	0.1306	0.2214	0.2122	0.2059
ForeDreamer	0.1770	0.2162	0.0414	0.0962	0.1573	0.1274	0.1742	0.1869	0.1471

Appendix

Table S9: Per-category data-leakage check on Prophet Arena for Qwen3.5-Flash. The table reports Brier score; lower values are better.

Method	Climate/Weather	Companies	Economics	Entertainment	Mentions	Other	Politics	Sports	Avg. ( $\downarrow$ )
No Information	0.2043	0.4272	0.1441	0.1973	0.2911	0.2310	0.2774	0.2210	0.2492
Full Text	0.2275	0.3719	0.0631	0.1424	0.2878	0.1410	0.2257	0.2075	0.2084
ForeDreamer	0.1811	0.3456	0.0506	0.1117	0.2704	0.1272	0.1927	0.1919	0.1839

Appendix

Table S10: Per-category data-leakage check on Prophet Arena for GPT-5.4-Nano. The table reports Brier score; lower values are better.

Dataset	Similarity metric	$\tau$	w/o tool reuse	w/ tool reuse	$\Delta$
FutureX	AST sequence	0.4	54.73%	68.18%	+13.46%
FutureX	AST sequence	0.5	67.66%	79.55%	+11.88%
FutureX	AST sequence	0.6	83.58%	84.09%	+0.51%
FutureX	Normalized sequence	0.4	23.88%	36.36%	+12.48%
FutureX	Normalized sequence	0.5	41.79%	45.45%	+3.66%
FutureX	Normalized sequence	0.6	57.21%	63.64%	+6.42%

Appendix

Table S11: Effect of compositional tool reuse on MemTool similarity clusters. Values are cluster ratios after converting to percentages; higher ratios indicate that fewer tools are merged into similarity clusters and therefore suggest less redundant tool generation.

\Delta

is computed as w/ tool reuse minus w/o tool reuse.

Figure S1: Hierarchical clustering of vanilla FutureX MemTools using normalized token-sequence similarity. The result shows substantial sequence-level overlap among generated tools, complementing the token-Jaccard analysis in the main text.

Figure S2: Hierarchical clustering of vanilla FutureX MemTools using AST node-sequence similarity. Even when lexical details are abstracted away, non-singleton clusters remain, indicating repeated structural patterns in generated tools.

Figure S3: Detailed pipeline-archetype analysis for FutureX MemGuides. Without diversity-guided exploration, all 10 MemGuides fall into Tool-Orchestrated Processing. With diversity-guided exploration, the guides cover four archetype patterns, showing broader procedural exploration.

Memory-Augmented Agents

Memory-augmented agents use stored experience and knowledge to act coherently across long tasks, changing environments, and repeated interactions.

See all 349 papers

ForecastCompass: Guiding Agentic Forecasting with Adaptive Factor Memory
May 29, 2026Yurui Chang, Yongkang Du, Yuanpu Cao +2Probabilistic Forecasting Memory-Augmented Agents
Pennsylvania State University
Agentic Time Machine as an Infrastructure for Future-Event Forecasting
Jun 19, 2026Jingyi Chai, Bingyang Zheng, Xiangrui Liu +5Multi-Agent LLM Systems LLM Agent Evaluation
Shanghai Jiao Tong University
Forecast-Dojo: Replayable Environments for Benchmarking and Training LLM Forecasting Agents
Sep 24, 2026Liqin Ye, Haorui Wang, Fardin Ahmed +8LLM Agent Evaluation Forecasting Benchmarks
Georgia Institute of Technology · Amazon · University of Florida

May 29, 2026J/K move · Enter open · S save

ForecastCompass: Guiding Agentic Forecasting with Adaptive Factor Memory

Yurui Chang, Yongkang Du, Yuanpu Cao, Jinghui Chen+1

Pennsylvania State University

Probabilistic Forecasting Memory-Augmented Agents

Open paper AskPDF arXiv

Agentic forecasting is important for decision-making in dynamic environments, but it remains challenging because agents must reason from incomplete, time-limited evidence and produce calibrated probabilities before outcomes are resolved. Memory provides a natural mechanism for transferring experience from resolved forecasts to future prediction tasks. However, existing agent-memory methods are not tailored to forecasting, as they typically store past interactions, reflections, or factual associations without explicitly representing reusable predictive factors or calibration knowledge. We propose ForecastCompass (FoCo), an adaptive factor-based memory framework for agentic forecasting. FoCo organizes forecasting experience with a hierarchical forecasting-task taxonomy, enabling retrieval task-relevant forecasting knowledge. It maintains two complementary memory components: factor memory, which captures reusable predictive dimensions, and reasoning memory, which encodes probability updating, uncertainty handling, and calibration principles. Using retrospective analyses as learning signals, FoCo iteratively revises memory through a verbalized memory-revision procedure, enabling the agent to accumulate transferable forecasting knowledge over time. Experiments on Prophet Arena and FutureX with GPT-5-mini and Gemini-2.5-Flash show that FoCo improves both probabilistic accuracy and calibration.

Abstract

Memory-Augmented Agents

Explore similar work

ForecastCompass: Guiding Agentic Forecasting with Adaptive Factor Memory

Agentic Time Machine as an Infrastructure for Future-Event Forecasting

Forecast-Dojo: Replayable Environments for Benchmarking and Training LLM Forecasting Agents