ForeDreamer: Self-Evolving Memory for Future Event Prediction

Abstract

Open-web future event prediction requires agents to distill reliable signals from noisy, redundant, and incomplete evidence. Existing retrieval and memory mechanisms either pass retrieved information directly to agents or rely on simple memory functions such as storing and reusing prior information, leaving them insufficient for open-web forecasting. We propose ForeDreamer, a self-evolving dual-agent framework that transforms raw web evidence into structured memory before prediction. ForeDreamer separates factual memory, a question-specific evidence state for the current forecast, from experiential memory, persistent agent experience accumulated across forecasting episodes. A main agent searches for evidence and produces forecasts, while a memory-processing subagent converts 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 its effectiveness.

Overview of the ForeDreamer forecasting workflow and dual-track self-evolution — ForeDreamer turns cutoff-aware web evidence into factual memory for each forecast and learns reusable textual and procedural experience from rollout feedback.

Motivation

Traditional agent-memory systems are commonly designed to store, retrieve, and reuse relatively clean interaction histories. Open-web forecasting presents a different challenge: useful evidence is sparse, time-sensitive, and scattered across heterogeneous sources, while search results may be outdated, distracting, redundant, or mutually conflicting.

ForeDreamer treats forecasting as an evidence-to-memory transformation problem. Raw results are processed into a coherent, question-specific evidence interface before the forecasting agent reasons over them.

Task mismatch between traditional agent memory and open-web future event prediction — Existing memory mechanisms fit clean, user-centered context better than noisy and conflicting open-web evidence.

ForeDreamer Framework

Question-specific

Factual Memory

A processed evidence artifact for the current forecast, constructed from temporally valid web observations and scoped to one question.

Persistent

Experiential Memory

Reusable forecasting and evidence-processing experience that accumulates across episodes and guides future behavior.

Dual-Agent Forecasting Workflow

The main agent consults an Experience Bank, plans multi-round searches, integrates processed factual memory, and produces the final probability or label. For every search result, a memory-processing subagent follows a MemGuide—an explicit evidence processing workflow—and invokes executable MemTools through an artifact workspace. This separation makes evidence processing explicit and inspectable instead of asking the forecasting model to interpret a long prompt of raw results.

Cutoff-Aware Forecasting

Each question is paired with a temporal cutoff. Web search is constrained to information available at prediction time, and all processed evidence is accumulated before the final forecast. The complete rollout—including interactions, artifacts, ground truth, and evaluation metrics—then becomes feedback for self-evolution.

Dual-Track Self-Evolution

Textual Track

Validated add, modify, or remove operations evolve the Experience Bank for search planning, evidence integration, and forecast calibration.

Procedural Track

Rollout feedback drives validated updates to a MemGuide tree and its executable MemTools, improving how factual memory is constructed.

Validation Gate

Candidate memories and procedures are admitted only when they pass interface checks and improve validation performance on the evolution pool.

Efficient and Diverse Procedural Evolution

Vanilla rollout-based expansion can repeatedly generate near-duplicate tools and over-concentrate on one successful guide family. Compositional Tool Reuse selects compatible existing MemTools and generates only missing operations. Diversity-Guided Exploration adds a scheduler-controlled path that proposes category-diverse candidates without requiring a new rollout, while retaining the same validation gate and guide-tree provenance.

Procedural evidence-processing evolution in ForeDreamer — Hybrid procedural evolution combines rollout-based expansion, compositional tool reuse, and diversity-guided exploration.

Experiments & Results

ForeDreamer is evaluated on open-ended future event prediction with Qwen3.5-Flash and GPT-5.4-Nano. Prophet Arena measures calibration with Brier score (lower is better), while FutureX measures prediction accuracy (higher is better). Comparisons include full-text and RAG settings as well as HippoRAG 2, Mem0, MemoryOS, A-MEM, LightMem, and LangMem.

0.1471 Prophet Arena · Qwen Brier ↓

0.1839 Prophet Arena · GPT Brier ↓

0.4108 FutureX · Qwen Accuracy ↑

0.3883 FutureX · GPT Accuracy ↑

Main Comparison

Benchmark / Backbone	Full Text	Best Existing Baseline	ForeDreamer
Prophet Arena · Qwen3.5-Flash ↓	0.2059	0.1761	0.1471
Prophet Arena · GPT-5.4-Nano ↓	0.2084	0.1997	0.1839
FutureX · Qwen3.5-Flash ↑	0.3298	0.3495	0.4108
FutureX · GPT-5.4-Nano ↑	0.2766	0.3567	0.3883

“Best Existing Baseline” reports the strongest result among the compared RAG and agent-memory baselines for each benchmark and backbone.

Dual-Track Experience Ablation

Method	Prophet Arena ↓	FutureX ↑
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

Removing either evolution track degrades performance, showing that textual forecasting experience and procedural evidence-processing experience provide complementary benefits.

Procedural-Evolution Optimization Ablation

Method	Prophet Arena ↓	FutureX ↑
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

Both compositional tool reuse and diversity-guided exploration contribute to the full system. The gains also remain consistent across alternative search providers, retrieval budgets, interaction limits, and context lengths.

Analysis

Compositional Tool Reuse Reduces Redundancy

Vanilla procedural evolution often rebuilds tools with overlapping source-code logic. Under the same token-Jaccard clustering criterion, compositional reuse produces a higher cluster ratio across similarity thresholds, indicating a less redundant MemTool set.

MemTool cluster ratio comparison with and without compositional tool reuse — Reusing compatible operations avoids spending the evolution budget on near-duplicate tools.

Diversity-Guided Exploration Broadens Strategies

Rollout-only expansion remains concentrated in one MemGuide pipeline archetype. The exploration path covers several distinct archetypes, allowing ForeDreamer to search beyond local variants of successful guide families.

MemGuide archetypes with and without diversity-guided exploration — Diversity-guided exploration yields a broader set of evidence-processing pipelines.

Citation

@article{zhong2026foredreamer,
  title={ForeDreamer: A Self-Evolving Dual-Agent Memory Architecture for Future Event Prediction},
  author={Zhong, Linhao and Du, Zongze and Wu, Linyu and Bo, Yu and Li, Hourong and Jing, Chenchen and Chen, Hao and Xi, Yuling and Shen, Chunhua},
  journal={arXiv preprint arXiv:2608.20920},
  year={2026}
}