Back to Details

evo-memory

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Evo-Memory

Evo-Memory

A persistent learning layer that accumulates research knowledge across ideation and experimentation cycles. Maintains two memory stores and implements three evolution mechanisms that feed learned patterns back into future research.
一个持久化学习层，可在创意构思与实验周期中积累研究知识。它维护两种记忆存储，并实现三种演化机制，将习得的模式反馈到未来的研究中。

When to Use This Skill

何时使用该技能

  • User has completed an 
    idea-tournament
    and needs to update Ideation Memory
  • User has completed (or failed) an 
    experiment-pipeline
    and needs to update memory
  • User is starting a new research cycle and wants to load prior knowledge
  • User asks about research memory, learned patterns, or cross-cycle knowledge
  • User mentions "evo-memory", "update memory", "what worked before", "research history", "evolution"
  • 用户已完成
    idea-tournament
    ，需要更新创意记忆M_I
  • 用户已完成（或失败）
    experiment-pipeline
    ，需要更新记忆
  • 用户启动新的研究周期，希望加载过往知识
  • 用户询问研究记忆、习得模式或跨周期知识相关问题
  • 用户提及“evo-memory”、“update memory”、“what worked before”、“research history”、“evolution”

The Learning Layer

学习层

Research is iterative. Each cycle — from ideation through experimentation — generates knowledge that should inform the next cycle. Without persistent memory, every new project starts from scratch, repeating mistakes and rediscovering patterns.
Evo-memory solves this by maintaining two structured memory stores and three evolution mechanisms that extract, classify, and inject knowledge across cycles.
研究是迭代的。从创意构思到实验的每个周期都会产生可指导下一个周期的知识。如果没有持久化记忆，每个新项目都要从零开始，重复犯错并重新发现已有模式。
Evo-memory通过维护两种结构化记忆存储和三种演化机制来解决这个问题，这些机制可跨周期提取、分类并注入知识。

Two Memory Stores

两种记忆存储

Ideation Memory (M_I)

创意记忆（M_I）

Location: 
/memory/ideation-memory.md
Records what you've learned about research DIRECTIONS — which areas are promising and which are dead ends.
Two sections:
SectionWhat It ContainsExample Entry
Feasible DirectionsDirections that showed promise in prior cycles"Contrastive learning for few-shot classification — confirmed feasible, top-3 in tournament cycle 2"
Unsuccessful DirectionsDirections that were tried and failed, with failure classification"Autoregressive generation for real-time video — fundamental failure: latency constraint incompatible with autoregressive decoding"
Each entry records: Direction name, one-sentence summary, evidence (which cycle, what results), classification (feasible / implementation failure / fundamental failure), date.
How it's used: 
idea-tournament
reads M_I at the start of Phase 1. The paper uses embedding-based retrieval with cosine similarity, selecting the top-k_I most similar items (k_I=2 in experiments). Feasible directions from prior cycles can seed new tree branches. Unsuccessful directions are used during pruning — fundamental failures are pruned; implementation failures may be retried.
See assets/ideation-memory-template.md for the template.
存储位置：
/memory/ideation-memory.md
记录关于研究方向的习得内容——哪些领域有前景，哪些是死胡同。
两个板块：
板块包含内容示例条目
可行方向在前序周期中显示出前景的研究方向“对比学习用于少样本分类——已确认可行，在第2周期竞赛中排名前三”
不可行方向已尝试且失败的方向，包含失败分类“自回归生成用于实时视频——根本性失败：延迟约束与自回归解码不兼容”
每个条目记录：方向名称、一句话摘要、证据（所属周期、结果）、分类（可行/实现类失败/根本性失败）、日期。
使用方式：
idea-tournament
在第1阶段开始时读取M_I。论文中使用基于嵌入的余弦相似度检索，选择最相似的前k_I个条目（实验中k_I=2）。前序周期的可行方向可作为新创意树的种子分支。不可行方向用于剪枝——根本性失败的方向会被剪枝，实现类失败的方向可重试。
查看assets/ideation-memory-template.md获取模板。

Experimentation Memory (M_E)

实验记忆（M_E）

Location: 
/memory/experiment-memory.md
Records what you've learned about research STRATEGIES — which technical approaches and configurations work in practice.
The paper defines M_E as storing "reusable data processing and model training strategies." ESE jointly summarizes (i) a data processing strategy and (ii) a model training strategy. We extend this with two additional practical sections (architecture and debugging) for comprehensive coverage.
Two core sections (from paper) + two practical extensions:
SectionSourceWhat It ContainsExample Entry
Data Processing StrategiesPaper (core)Preprocessing, augmentation, and data handling patterns"For noisy sensor data: median filter before normalization reduces training instability by ~40%"
Model Training StrategiesPaper (core)Hyperparameters, training tricks, and training schedules"Learning rate warmup for 10% of steps prevents early divergence in transformer fine-tuning"
Architecture StrategiesExtensionDesign choices, module configurations, and structural patterns"Residual connections are critical for modules inserted deeper than 10 layers in transformers"
Debugging StrategiesExtensionDiagnostic patterns that resolved experiment failures"When loss plateaus after 50% of training: check gradient norm — clipping threshold may be too aggressive"
Each entry records: Strategy name, context (when to use this), evidence (which cycle, what results), generality (domain-specific or broadly applicable), date.
How it's used: 
experiment-pipeline
reads M_E at the start of each cycle. The paper uses embedding-based retrieval with cosine similarity, selecting the top-k_E most similar items (k_E=1 in experiments). Relevant strategies from prior cycles inform hyperparameter choices, data processing decisions, and debugging approaches, reducing the number of attempts needed.
See assets/experiment-memory-template.md for the template.
存储位置：
/memory/experiment-memory.md
记录关于研究策略的习得内容——哪些技术方法和配置在实践中有效。
论文中定义M_E用于存储“可复用的数据处理和模型训练策略”。ESE联合总结（i）数据处理策略和（ii）模型训练策略。我们扩展了两个实用板块（架构与调试）以实现全面覆盖。
两个核心板块（来自论文）+两个实用扩展板块：
板块来源包含内容示例条目
数据处理策略论文（核心）预处理、数据增强和数据处理模式“针对含噪传感器数据：归一化前使用中值滤波器可将训练不稳定性降低约40%”
模型训练策略论文（核心）超参数、训练技巧和训练调度“学习率预热10%的步骤可防止Transformer微调中的早期发散”
架构策略扩展设计选择、模块配置和结构模式“残差连接对于插入Transformer第10层之后的模块至关重要”
调试策略扩展解决实验失败的诊断模式“当训练进行到50%后损失停滞：检查梯度范数——裁剪阈值可能过于严格”
每个条目记录：策略名称、适用场景（何时使用）、证据（所属周期、结果）、通用性（领域特定或广泛适用）、日期。
使用方式：
experiment-pipeline
在每个周期开始时读取M_E。论文中使用基于嵌入的余弦相似度检索，选择最相似的前k_E个条目（实验中k_E=1）。前序周期的相关策略可为超参数选择、数据处理决策和调试方法提供参考，减少尝试次数。
查看assets/experiment-memory-template.md获取模板。

Three Evolution Mechanisms

三种演化机制

IDE — Idea Direction Evolution

IDE — 创意方向演化

Trigger: After 
idea-tournament
completes (Phase 3 direction summary is available).
Purpose: Extract promising research directions from the tournament results and store them in M_I for future cycles.
Paper Prompt: Use the IDE prompt from references/paper-prompts.md as the primary extraction mechanism. Fill in 
{user_goal}
from the original research direction and 
{top_ranked_ideas}
from 
/direction-summary.md
, then reason through the prompt step by step. The output (DIRECTION SUMMARY with Title, Core idea, Why promising, Requirements, Validation plan) feeds directly into the steps below.
Process:
  1. Read current M_I from 
    /memory/ideation-memory.md
  2. Run the paper's IDE prompt (see above), reasoning through it step by step
  3. For each direction in the prompt output, abstract it to a reusable level. "Attention-based feature selection for 3D point clouds" becomes "Cross-domain attention mechanisms for sparse data" — specific enough to be useful, abstract enough to transfer.
  4. Check M_I for existing entries on similar directions. Update if exists, append if new.
  5. If any previously "feasible" direction was found to be exhausted during this cycle, update its status.
  6. Write an evolution report documenting what changed and why.
Key principle: Store directions, not ideas. A direction like "contrastive learning for structured data" can spawn many specific ideas across future cycles. A specific idea like "SimCLR with graph augmentations on molecular datasets" is too narrow to be reusable.
See references/ide-protocol.md for the full process.
触发条件：
idea-tournament
完成后（第3阶段方向摘要已生成）。
目的：从竞赛结果中提取有前景的研究方向并存储到M_I，供未来周期使用。
论文提示词：使用references/paper-prompts.md中的IDE提示词作为主要提取机制。将
{user_goal}
替换为原始研究方向，
{top_ranked_ideas}
替换为
/direction-summary.md
中的内容，然后逐步执行提示词。输出内容（包含标题、核心创意、优势、要求、验证方案的方向摘要）直接用于以下步骤。
流程：
  1. 从
    /memory/ideation-memory.md
    读取当前M_I
  2. 运行论文中的IDE提示词（如上所述），逐步推理
  3. 对提示词输出中的每个方向进行抽象，达到可复用的级别。例如，“基于注意力的3D点云特征选择”抽象为“跨域注意力机制用于稀疏数据”——具体到足以实用，同时抽象到可迁移。
  4. 检查M_I中是否存在相似方向的现有条目。若存在则更新，不存在则追加。
  5. 如果本周期中发现任何之前标记为“可行”的方向已被穷尽，更新其状态。
  6. 撰写演化报告，记录变更内容及原因。
核心原则：存储研究方向，而非具体创意。像“对比学习用于结构化数据”这样的方向可在未来周期中衍生出许多具体创意；而“在分子数据集上使用带图增强的SimCLR”这样的具体创意过于狭窄，无法复用。
查看references/ide-protocol.md获取完整流程。

IVE — Idea Validation Evolution

IVE — 创意验证演化

Trigger (two conditions, following the paper):
  1. Rule-based: The engineer cannot find any executable code within the pre-defined budget at any stage — the code simply doesn't run.
  2. LLM-based: Experiments complete but the proposed method performs worse than the baseline, as determined by analyzing the execution report W.
Purpose: Classify WHY the method failed and update M_I accordingly. This is the most critical evolution mechanism because it prevents future cycles from repeating dead-end directions.
Paper Prompt: Use the IVE prompt from references/paper-prompts.md as the primary classification mechanism. Fill in 
{research_proposal}
from 
/research-proposal.md
and 
{execution_report}
from the stage trajectory logs, then reason through the prompt step by step. The prompt classifies the failure as FAILED(NoExecutableWithinBudget), FAILED(WorseThanBaseline), or NOT_FAILED.
After running the paper prompt:
  • FAILED(NoExecutableWithinBudget) → Implementation failure (retryable). Record as "retry with fixes" in M_I.
  • FAILED(WorseThanBaseline) → Use the 5-question diagnostic below to distinguish implementation vs fundamental failure.
  • NOT_FAILED → No IVE update needed.
Five diagnostic questions (for WorseThanBaseline cases):
  1. Did any variant show partial success? (Yes → implementation failure)
  2. Does the hypothesis hold for simpler problems? (No → fundamental failure)
  3. Have related approaches succeeded in published work? (Yes → implementation failure)
  4. Were failure patterns consistent across implementations? (Yes → fundamental failure)
  5. Can you identify specific bugs in trajectory logs? (Yes → implementation failure)
If 3+ answers point to one type, classify as that type. If split, classify as implementation failure (more conservative — allows retry).
Retry escalation rule: If a direction has been classified as "implementation failure" 3 times across different cycles, escalate to a careful re-evaluation — three separate implementation failures may indicate the direction is harder than it appears. Consider reclassifying as fundamental.
See references/ive-protocol.md for the full process and worked examples.
触发条件（符合论文的两个条件）：
  1. 基于规则：工程师在预定义预算内的任何阶段都无法找到可执行代码——代码根本无法运行。
  2. 基于大语言模型：实验完成，但分析执行报告W后发现，所提出的方法性能不如基线模型。
目的：分类方法失败的原因并相应更新M_I。这是最关键的演化机制，因为它可防止未来周期重复尝试死胡同方向。
论文提示词：使用references/paper-prompts.md中的IVE提示词作为主要分类机制。将
{research_proposal}
替换为
/research-proposal.md
中的内容，
{execution_report}
替换为阶段轨迹日志中的内容，然后逐步执行提示词。提示词会将失败分类为FAILED(NoExecutableWithinBudget)、FAILED(WorseThanBaseline)或NOT_FAILED。
运行论文提示词后：
  • FAILED(NoExecutableWithinBudget) → 实现类失败（可重试）。在M_I中记录为“修复后重试”。
  • FAILED(WorseThanBaseline) → 使用以下5个诊断问题区分实现类与根本性失败。
  • NOT_FAILED → 无需更新IVE。
五个诊断问题（针对性能不如基线的情况）：
  1. 是否有变体显示出部分成功？（是→实现类失败）
  2. 假设在更简单的问题上是否成立？（否→根本性失败）
  3. 相关方法在已发表的研究中是否成功？（是→实现类失败）
  4. 失败模式在不同实现中是否一致？（是→根本性失败）
  5. 能否在轨迹日志中识别出具体bug？（是→实现类失败）
若3个及以上答案指向某一类型，则分类为该类型。若答案分歧，则分类为实现类失败（更保守——允许重试）。
重试升级规则：如果某个方向在不同周期中被3次分类为“实现类失败”，则需进行仔细的重新评估——三次独立的实现类失败可能表明该方向比预期更难。考虑将其重新分类为根本性失败。
查看references/ive-protocol.md获取完整流程及示例。

ESE — Experiment Strategy Evolution

ESE — 实验策略演化

Trigger: After 
experiment-pipeline
succeeds — all 4 stages complete and gates met.
Purpose: Distill reusable strategies from the successful experiment run and store them in M_E for future cycles.
Paper Prompt: Use the ESE prompt from references/paper-prompts.md as the primary extraction mechanism. Fill in 
{research_proposal}
from 
/research-proposal.md
and 
{trajectories}
from all 4 stage trajectory logs, then reason through the prompt step by step. The prompt outputs DATA SUMMARY and MODEL SUMMARY, which map to our Data Processing Strategies and Model Training Strategies sections.
Process:
  1. Run the paper's ESE prompt (see above), reasoning through it step by step
  2. Use the DATA SUMMARY output to populate the Data Processing Strategies section of M_E
  3. Use the MODEL SUMMARY output to populate the Model Training Strategies section of M_E
  4. After the prompt run, manually extract from trajectory logs:
    • Architecture decisions (extension): Which design choices were key to performance?
    • Debugging patterns (extension): Which diagnostic approaches resolved failures fastest?
  5. For each identified pattern, assess generality:
    • Is this domain-specific (only works for this type of data/model)?
    • Or broadly applicable (likely to work in other contexts)?
  6. Check M_E for existing similar entries. Update if exists, append if new.
  7. Write an evolution report documenting the extracted strategies.
Generalization guidelines: A strategy is broadly applicable if it addresses a general challenge (training instability, overfitting, slow convergence) rather than a domain-specific characteristic. When in doubt, record the context alongside the strategy and let future users judge applicability.
See references/ese-protocol.md for the full process.
触发条件：
experiment-pipeline
成功完成——所有4个阶段均完成并通过关卡。
目的：从成功的实验运行中提炼可复用策略并存储到M_E，供未来周期使用。
论文提示词：使用references/paper-prompts.md中的ESE提示词作为主要提取机制。将
{research_proposal}
替换为
/research-proposal.md
中的内容，
{trajectories}
替换为所有4个阶段的轨迹日志，然后逐步执行提示词。提示词输出的DATA SUMMARY和MODEL SUMMARY对应我们的数据处理策略和模型训练策略板块。
流程：
  1. 运行论文中的ESE提示词（如上所述），逐步推理
  2. 使用DATA SUMMARY输出填充M_E的数据处理策略板块
  3. 使用MODEL SUMMARY输出填充M_E的模型训练策略板块
  4. 运行提示词后，从轨迹日志中手动提取：
    • 架构决策（扩展）：哪些设计选择是性能关键？
    • 调试模式（扩展）：哪些诊断方法最快解决了失败问题？
  5. 对每个识别出的模式评估通用性：
    • 是领域特定的（仅适用于此类数据/模型）？
    • 还是广泛适用的（可能在其他场景中有效）？
  6. 检查M_E中是否存在相似的现有条目。若存在则更新，不存在则追加。
  7. 撰写演化报告，记录提取的策略。
通用性指南：如果策略解决的是通用挑战（训练不稳定性、过拟合、收敛缓慢）而非领域特定特征，则属于广泛适用。若不确定，请在记录策略的同时记录其适用场景，由未来用户判断是否适用。
查看references/ese-protocol.md获取完整流程。

Reading Memory at Cycle Start

周期开始时读取记忆

When starting a new research cycle (loading 
idea-tournament
or 
experiment-pipeline
):
  1. Read 
    /memory/ideation-memory.md
    and 
    /memory/experiment-memory.md
  2. Summarize relevant entries to inject into the current context
  3. For 
    idea-tournament
    : Use M_I feasible directions to seed tree branches. Use M_I unsuccessful directions (fundamental failures only) during pruning.
  4. For 
    experiment-pipeline
    : Use M_E strategies to inform hyperparameter ranges, training schedules, and debugging approaches.
Don't blindly apply old strategies. Context matters. A strategy that worked for image classification may not work for text generation. Always check the recorded context against the current problem.
Retrieval method: The paper uses embedding-based cosine similarity for retrieval. In practice, perform this semantic comparison by reading each entry's Summary/Context and Retrieval Tags, then judging relevance to the current goal. If automated embedding tools are available in your environment, use those instead for larger memory stores.
当启动新的研究周期（加载
idea-tournament
或
experiment-pipeline
）时：
  1. 读取
    /memory/ideation-memory.md
    和
    /memory/experiment-memory.md
  2. 总结相关条目并注入当前上下文
  3. 对于
    idea-tournament
    ：使用M_I中的可行方向作为创意树的种子分支。使用M_I中的不可行方向（仅根本性失败）进行剪枝。
  4. 对于
    experiment-pipeline
    ：使用M_E中的策略指导超参数范围、训练调度和调试方法。
不要盲目应用旧策略。上下文很重要。在图像分类中有效的策略可能不适用于文本生成。始终检查记录的上下文是否与当前问题匹配。
检索方法：论文中使用基于嵌入的余弦相似度检索。在实践中，可通过读取每个条目的摘要/上下文和检索标签，判断其与当前目标的语义相关性。若环境中有自动化嵌入工具，对于较大的记忆存储可使用该工具。

For idea-tournament (inject M_I)

针对idea-tournament（注入M_I）

  1. Read 
    /memory/ideation-memory.md
  2. Select the top-k_I=2 entries most relevant to the user's current goal. Compare the user's goal statement against each entry's Summary and Retrieval Tags for semantic similarity.
  3. For each selected feasible direction: incorporate it as a seed branch at Level 1 of the idea tree. Example injection: "Prior cycle found 'Modality-aware model compression' promising (Elo 1548, cycle 3). Use as a Level 1 branch alongside new technique variants."
  4. For each unsuccessful direction with 
    Failure Classification: Fundamental
    : flag for pruning. Example injection: "Prior cycle confirmed 'Autoregressive real-time video generation' is a fundamental failure (O(n) latency). Prune any tree branch matching this pattern."
  1. 读取
    /memory/ideation-memory.md
  2. 选择与用户当前目标最相关的前k_I=2个条目。将用户的目标陈述与每个条目的摘要和检索标签进行语义相似度比较。
  3. 对于每个选中的可行方向：将其作为创意树第1层的种子分支。示例注入：“前序周期发现‘模态感知模型压缩’有前景（Elo 1548，第3周期）。将其作为第1层分支，与新技术变体并列。”
  4. 对于每个标记为“失败分类：根本性失败”的不可行方向：标记为待剪枝。示例注入：“前序周期确认‘自回归实时视频生成’为根本性失败（O(n)延迟）。剪枝所有匹配该模式的创意树分支。”

For experiment-pipeline (inject M_E)

针对experiment-pipeline（注入M_E）

  1. Read 
    /memory/experiment-memory.md
  2. Select the top-k_E=1 entry most relevant to the current experiment domain. Compare the experiment's problem description against each entry's Context and Category.
  3. Inject the selected strategy as context for all stages. Example injection: "Prior cycle found 'Cosine annealing with warm restarts (T_0=10, T_mult=2)' effective for transformer fine-tuning on small datasets (confirmed, 2 cycles). Apply in Stage 2 tuning as the default schedule."
  4. Also scan the Debugging Strategies section for any entries matching the current domain — these can save significant time when diagnosing failures.
  1. 读取
    /memory/experiment-memory.md
  2. 选择与当前实验领域最相关的前k_E=1个条目。将实验的问题描述与每个条目的上下文和类别进行比较。
  3. 将选中的策略作为所有阶段的上下文注入。示例注入：“前序周期发现‘带热重启的余弦退火（T_0=10，T_mult=2）’在小数据集上的Transformer微调中有效（已确认，2个周期）。在第2阶段调优中作为默认调度使用。”
  4. 同时扫描调试策略板块，查找与当前领域匹配的条目——这些条目可在诊断失败时节省大量时间。

Memory Maintenance

记忆维护

Pruning Stale Entries

剪枝陈旧条目

Periodically review both memory stores and remove or archive entries that are no longer relevant:
  • Entries older than 10 cycles without being referenced
  • Strategies superseded by strictly better alternatives
  • Directions in fields that have fundamentally shifted (new paradigms, new state-of-the-art)
定期检查两种记忆存储，删除或归档不再相关的条目：
  • 超过10个周期未被引用的条目
  • 被更优方案取代的策略
  • 所属领域已发生根本性转变的方向（新范式、新SOTA）

Version Tracking

版本跟踪

Each memory file maintains a 
Last Updated
field and a cycle counter. When entries are modified (not just appended), note what changed in the evolution report. This creates an audit trail of how your research knowledge evolves.
每个记忆文件都维护“最后更新”字段和周期计数器。当条目被修改（不仅仅是追加）时，在演化报告中记录变更内容。这会创建研究知识演化的审计轨迹。

Evolution Reports

演化报告

After each evolution mechanism triggers, generate a report saved to 
/memory/evolution-reports/cycle_N_type.md
:
  • What changed (added, updated, or removed entries)
  • Why (evidence from the triggering cycle)
  • Expected impact on future cycles
See assets/evolution-report-template.md for the template.
每次触发演化机制后，生成报告并保存到
/memory/evolution-reports/cycle_N_type.md
：
  • 变更内容（添加、更新或删除的条目）
  • 变更原因（触发周期的证据）
  • 对未来周期的预期影响
查看assets/evolution-report-template.md获取模板。

Counterintuitive Memory Rules

反直觉记忆规则

Prioritize these rules when updating and using memory:
  1. Abstract before storing: Store directions and strategies, not specific experiment details. "Contrastive learning improves few-shot classification" is reusable across many projects; "set lr=0.001 for ResNet-50 on CIFAR-10" is not. The goal is transferable knowledge, not a lab notebook.
  2. Failed directions are more valuable than successful ones: Knowing what NOT to try saves more time than knowing what worked. Success stories are published in papers — everyone can access them. Failure stories are rarely shared, making your failure memory a unique competitive advantage.
  3. Implementation failures are not direction failures: The most common evolution mistake is marking a good direction as failed because the implementation was buggy. IVE exists specifically to make this distinction. When in doubt, classify as implementation failure — it's cheaper to retry a good idea than to permanently discard it.
  4. Memory decays without pruning: A strategy that worked 10 cycles ago on different data may no longer be relevant. Accumulating stale entries adds noise that makes it harder to find useful strategies. Prune actively — a smaller, curated memory is more valuable than a large, noisy one.
  5. Cross-pollination beats deep specialization: Strategies from M_E in one domain often transfer to another. Learning rate warmup helps in NLP AND vision AND speech. Review the full M_E before starting a new experiment pipeline, not just domain-specific entries.
  6. The evolution report is for humans: Write reports that a researcher — not just an AI agent — can understand and act on. Include enough context that someone reading the report 6 months later understands WHY the change was made, not just WHAT changed.
更新和使用记忆时，请优先遵循以下规则：
  1. 先抽象再存储：存储方向和策略，而非具体实验细节。“对比学习提升少样本分类性能”可在多个项目中复用；“在CIFAR-10上为ResNet-50设置lr=0.001”则不可复用。目标是可迁移知识，而非实验记录本。
  2. 失败方向比成功方向更有价值：知道什么不该尝试比知道什么有效更节省时间。成功案例会发表在论文中——所有人都能获取。失败案例很少被分享，因此你的失败记忆是独特的竞争优势。
  3. 实现类失败不等于方向失败：最常见的演化错误是因为实现存在bug，就将一个好的方向标记为失败。IVE机制的存在正是为了区分这两者。若不确定，分类为实现类失败——重试一个好创意的成本远低于永久丢弃它。
  4. 不剪枝的记忆会衰退：10个周期前在不同数据上有效的策略可能已不再相关。积累陈旧条目会增加噪声，使查找有用策略变得更困难。积极剪枝——小而精的记忆比大而杂的记忆更有价值。
  5. 跨领域复用优于深度专精：M_E中某一领域的策略通常可迁移到另一领域。学习率预热在NLP、计算机视觉和语音领域都有效。启动新实验流程前，请查看完整的M_E，而非仅查看领域特定的条目。
  6. 演化报告是为人类撰写的：撰写的报告应让研究人员（而非仅AI Agent）能够理解并采取行动。包含足够的上下文，让6个月后阅读报告的人能够理解变更的原因，而非仅仅知道变更内容。

Memory Integration Points

记忆集成点

How evo-memory connects to other skills in the pipeline:
TriggerSource SkillMechanismMemory Updated
Tournament completed
idea-tournament
IDEM_I (feasible directions)
No executable code within budget, or method underperforms baseline
experiment-pipeline
IVEM_I (unsuccessful directions)
Pipeline succeeded
experiment-pipeline
ESEM_E (data processing + model training; optionally architecture + debugging)
New cycle starts
idea-tournament
Read (top-k_I=2)M_I read for seeding/pruning
New cycle starts
experiment-pipeline
Read (top-k_E=1)M_E read for strategy guidance
Evo-memory与流程中其他技能的连接方式：
触发条件来源技能机制更新的记忆
竞赛完成
idea-tournament
IDEM_I（可行方向）
无可用可执行代码，或方法性能低于基线
experiment-pipeline
IVEM_I（不可行方向）
流程成功
experiment-pipeline
ESEM_E（数据处理+模型训练；可选架构+调试）
新周期启动
idea-tournament
读取（前k_I=2）读取M_I用于种子分支/剪枝
新周期启动
experiment-pipeline
读取（前k_E=1）读取M_E用于策略指导

Reference Navigation

参考导航

TopicReference FileWhen to Use
IDE process detailside-protocol.mdAfter completing idea-tournament
IVE process detailsive-protocol.mdAfter experiment-pipeline failure (no executable code or method underperforms)
ESE process detailsese-protocol.mdAfter experiment-pipeline succeeds
Paper's actual promptspaper-prompts.mdReference for exact IDE/IVE/ESE prompt design
Memory data structuresmemory-schema.mdUnderstanding M_I and M_E formats
Ideation memory templateideation-memory-template.mdInitializing M_I
Experiment memory templateexperiment-memory-template.mdInitializing M_E
Evolution report templateevolution-report-template.mdDocumenting memory updates
主题参考文件使用场景
IDE流程细节ide-protocol.md完成idea-tournament后
IVE流程细节ive-protocol.mdexperiment-pipeline失败后（无可用可执行代码或方法性能低于基线）
ESE流程细节ese-protocol.mdexperiment-pipeline成功后
论文原始提示词paper-prompts.md参考IDE/IVE/ESE提示词的精确设计
记忆数据结构memory-schema.md理解M_I和M_E的格式
创意记忆模板ideation-memory-template.md初始化M_I
实验记忆模板experiment-memory-template.md初始化M_E
演化报告模板evolution-report-template.md记录记忆更新