为什么需要会进化的记忆系统 🍅
传统 Agent 记忆立在一个核心假设上：记忆结构预先设计好，Agent 只在这个固定框架里存取、检索、调度信息。任务简单时这个假设成立；任务一复杂，它就会以可预期的方式失效。记忆规模上来之后，近似重复的经验不断累积，冗余开始干扰检索和决策，而不是支撑它们。单一的静态架构也照顾不了异构的工作负载：不同任务形态本来就需要不同的记忆 pipeline，而设计者冻结下来的那个结构，隐含地只对他设计时心里想的那种任务分布调过优。
自演进记忆是第一步回应：让 Agent 在运行中自适应调整记忆策略。但此前的工作把进化停在了策略层，也就是“怎么用记忆”，底下的架构依然固定。MemEvolve 把边界再往下推了一层：让记忆系统本身进入可进化空间，把记忆结构当成可优化的对象，Agent 由此获得结构级的适应能力。
MemEvolve 工作详解 🍅
与把记忆机制当固定模块的传统做法不同，MemEvolve 的核心思想是把记忆内容和记忆系统架构一起纳入可进化的搜索空间。
实验结果显示，在多任务推理和长程依赖场景中，MemEvolve 稳定拿到更强的任务表现。我对这个结果的读法是：把可进化范围从策略层拓宽到结构层，抬高的是系统能力的天花板。固定架构下可用的策略被那个架构框死，把架构放开，这个上界也就跟着解除了。
 |
|---|
| 图 1. MemEvolve 与多种主流自演进记忆系统在 xBench-DS、WebWalkerQA、GAIA 上的对比结果；底座为 Flash-Searcher + GPT-5-Mini |
MemEvolve 把进化对象扩展到两个层级：
-
记忆本身的进化。 这一层管的是记忆如何被使用、如何随时间成形。记忆被拆解成一组可进化的功能单元：编码方式、存储策略、检索机制、管理规则。进化过程在任务级奖励的驱动下对这些组件做选择性优化，让机制本身逐步适配不同推理任务的需求。
-
记忆系统架构的进化。 第二层管的是这些能力如何组织成一个完整、可扩展的结构。在 EvolveLab 提供的统一抽象下，进化不仅能调整单个模块的内部行为，还能重构模块之间的连接关系、调用顺序和层级组织。差异化的记忆模块由进化过程自己长出来，而不是人手设计出来。
 |
|---|
| 图 2. MemEvolve 整体框架示意图 |
进化出的记忆系统实例 🍅
MemEvolve 进化出的三个记忆系统，与其当成品看，不如当证据看。不同的选择压力下，同一套进化过程落在了设计空间里明显不同的点上；结构进化要想不沦为口号，需要证明的恰恰就是这件事。
| 记忆系统 | 适用场景 | 核心优势 | 主要代价 |
|---|
| RIVA | 多跳推理、规划任务、推理轨迹复用 | 推理过程具备高度可重组性 | 记忆操作复杂，计算与存储开销较高 |
| CEREBRA | 跨任务知识积累、长期稳定决策 | 任务反馈驱动下自发形成分层知识结构 | 更新节奏缓慢，结构惯性强，新知识接入成本高 |
| LIGHTWEIGHT | 资源受限环境、短时任务 | 极简记忆策略带来高效推理 | 缺乏长期知识沉淀与跨任务迁移能力，容易退化为短期决策机制 |
RIVA：以推理可复用性为核心的记忆流 🍅
很多任务不是回忆题。它们要求 Agent 在多个时间步、多个中间状态之间反复引用、重组、修正已有知识；对这一类工作来说，值得记住的是推理，不是原始信息。RIVA 押的正是这一点：它存的不是原始内容，而是半结构化的推理片段，日后可以当作可组合的推理资产重新拼装。
它的检索也不止于相似度匹配。RIVA 会优先激活那些历史上参与过成功推理轨迹的片段，这是一种经验驱动的偏置：给记忆称重的是被证明过的有用性，而不是表面相关性。系统层面，记忆和推理之间高频双向互动：中间推理结果实时写回记忆，已有条目可以被选择性覆盖或更新，整个记忆空间是持续被修整的，不是只增不改的。这套设计的代价在上面那张表里写得清楚：可重组的推理需要复杂的记忆操作，计算和存储开销随之上涨。
CEREBRA：分层且可治理的长程记忆 🍅
CEREBRA 的进化路径指向另一头：长期知识的稳定性与可治理性。它最鲜明的特征是自发长出的分层记忆结构，不同层级承担不同的语义角色，层与层之间的信息流动受隐含的门控机制约束。
要紧的是这套治理从哪来。不是手写规则，也不是预设 schema，而是被任务级反馈信号一点点塑造出来的。在奖励驱动的迭代选择中，系统进化出一个层级结构，很像知识随时间沉淀、固结的过程。也正是这种靠反馈固结出来的稳定，让它改起来慢：更强的结构惯性、更高的新知识接入成本，是长程可靠性的价钱。
LIGHTWEIGHT：资源约束下的极简记忆 🍅
当进化目标里加进了显式或隐式的资源约束，计算和存储开销上的任何劣势都会被选择放大。LIGHTWEIGHT 就是这种压力下长出来的记忆形态。它主动放弃长期知识积累和推理轨迹复用，把记忆收窄成短生命周期、任务相关的缓存。只有系统判断当前信息对后续决策价值重大时才写入，检索策略也刻意从简。这套设计把记忆模块对整体推理的干扰压到最低，换来资源受限环境下的稳定表现；表里也写了代价：任务一拉长，它就容易退化成短期决策机制。
MemEvolve：从被设计的记忆模块到可进化的结构 🍅
MemEvolve 打破了记忆结构必须人为固定的前提，把记忆从预先设计的组件提升为可以被学习和进化的对象。
当前 Agent 平台的记忆架构多是设计驱动，而非演化驱动。借鉴 MemEvolve，平台的记忆结构可以转向任务特性导向：让结构从任务需求里长出来，而不是对着一个假想的负载预先定死。
启示一：引入 EvolveLab 四模块抽象，构建可搜索的结构空间 🍅
提升 Agent 能力不必然依赖更大的模型或更长的上下文窗口。结构本身就能承载能力。 当记忆的编码方式、存储策略、检索路径、模块连接关系都可以被进化搜索时，性能增益就不再只来自参数扩张，而开始来自结构层面的重组与适配。
过去 Agent 平台常见的能力增长路径是一条线性扩张：模型升级 → RAG 增强 → GraphRAG → 工具生态扩展 → context 扩展。这条路每一步加的都是容量，没有一步加的是结构适应性；后者才是接下来值得补的维度。
MemEvolve 的工程底座 EvolveLab 给了一个可参考的做法：把标准化的记忆模块解耦成独立、可替换、可进化的单元，把记忆架构设计转化为模块配置的组合搜索问题。
对 Agent 平台而言，照 EvolveLab 的思路给记忆系统做模块化解耦，还没开始搜索就已经有收益：单个模块可以独立替换、独立测试，不必重写整个系统。这层解耦也是后续结构搜索和自动调优的前提，它让结构自适应成为一条可增量推进的工程路线，而不是一场推倒重来。
启示二：从固定的记忆范式到可选择、可演化的记忆空间 🍅
我从 MemEvolve 里读到的工程教训是：在系统设计时押死一条“正确的”记忆 pipeline，是一步不必要的先手亏。完全可以构建一个可搜索的记忆空间，让不同的任务分布在运行中逐步塑造适合自己的结构。MemEvolve 的外层循环展示了这在实践里需要什么，一套“诊断 + 设计”的工作流：先用 Pareto 多目标排序（任务成功率、成本、延迟）选出候选架构；再分析失败案例和记忆访问模式，生成缺陷画像；最后基于这份诊断信号定向合成改进后的架构。Pareto 这一步不是花架子：记忆设计没有单一最优解，成功率、成本、延迟互相牵制，只按一个轴排序，恰恰会把搜索最需要的多样性折叠掉。
对 Agent 平台而言，原材料现成就有：任务日志里躺着失败轨迹、推理断点、记忆访问统计，足够拼出任务级的缺陷画像。这些信号可以驱动推理型记忆、层级知识记忆、轻量缓存策略三条线的定向迭代。比起一次性的人工设计决策，这样的循环更系统，也更配得上一个每天见过无数种任务分布的 Agent 平台。
这个转变把 Agent 记忆研究推向自组织、自优化的范式，也更接近我们期待智能系统长出的那种内生知识结构。
Why There Is a Need for Evolving Memory Systems 🍅
Traditional agent memory rests on one core assumption: the memory structure is designed in advance, and the agent merely accesses, retrieves, and schedules information inside that fixed frame. The assumption held while tasks were simple, and it degrades in predictable ways once they are not. As memory scales, near-duplicate experiences accumulate, and the redundancy starts interfering with retrieval and decision-making rather than supporting it. A single static architecture also cannot serve heterogeneous workloads: different task forms genuinely demand different memory pipelines, and whatever structure the designer froze in place was tuned, implicitly, to the task distribution they had in mind at design time.
Self-evolving memory was the first response: let the agent adapt its memory strategies during execution. But earlier work kept evolution at the strategy level (how to use the memory) while the architecture underneath stayed fixed. MemEvolve pushes the boundary one level down. It brings the memory system itself into the evolvable space, treating memory structure as an optimizable object, and thereby gives agents adaptability at the structural level.
Technical Breakdown of MemEvolve 🍅
Where traditional approaches treat the memory mechanism as a fixed module, MemEvolve’s core idea is to place both memory content and the memory system architecture itself into an evolvable search space.
Experimental results show that in multi-task reasoning and long-range dependency scenarios, MemEvolve consistently achieves stronger task performance. The reading I take from this: widening the evolvable scope from strategy-level adaptation to structural optimization raises the ceiling of what the system can reach: the strategies available to a fixed architecture are bounded by that architecture, and relaxing the architecture removes the bound.
|
|---|
| Figure 1. Comparison of MemEvolve with Mainstream Self-Evolving Memory Systems on xBench-DS, WebWalkerQA, and GAIA. Backbone is Flash-Searcher + GPT-5-Mini |
MemEvolve extends the scope of evolution across two hierarchical levels:
-
Evolution of memory itself. This layer governs how memory is used and shaped over time. Memory is decomposed into evolvable functional units (encoding schemes, storage strategies, retrieval mechanisms, and management rules), and the evolutionary process, driven by task-level rewards, selectively optimizes these components so the mechanisms gradually adapt to the demands of different reasoning tasks.
-
Evolution of the memory system architecture. The second layer governs how those capabilities are organized into a coherent, scalable structure. Under the unified abstraction provided by EvolveLab, evolution can adjust not only the internal behavior of individual modules but also inter-module connections, invocation order, and hierarchical organization. Differentiated memory modules emerge from the evolutionary process itself rather than from manual design.
|
|---|
| Figure 2. MemEvolve General Structure Diagram |
Evolved Memory System Instances 🍅
The three memory systems MemEvolve evolved are worth reading less as products than as evidence. Under different selection pressures, the same evolutionary process settles on visibly different points of the design space, which is exactly what structural evolution needs to demonstrate to be more than a slogan.
| Memory System | Applicable Scenarios | Core Strengths | Primary Trade-offs |
|---|
| RIVA | Multi-hop reasoning, planning tasks, reasoning trace reuse | Highly recomposable reasoning processes | Complex memory operations with higher computational and storage overhead |
| CEREBRA | Cross-task knowledge accumulation, long-term stable decision-making | Formation of hierarchical knowledge structures driven by task feedback | Slower update dynamics, stronger structural inertia, higher adaptation cost to newly introduced knowledge |
| LIGHTWEIGHT | Resource-constrained environments, short-horizon tasks | Efficient reasoning through minimalist memory strategies | Limited long-term knowledge retention and cross-task transfer, prone to degenerating into short-term decision mechanisms |
RIVA: A Memory Flow Centered on Reasoning Reusability 🍅
Many tasks are not recall problems. They require the agent to repeatedly reference, reorganize, and revise existing knowledge across multiple time steps and intermediate states, and for that class of work, what deserves to be remembered is the reasoning, not the raw information. RIVA leans into exactly this: instead of storing raw content, it retains semi-structured reasoning fragments that can later be recombined as composable reasoning assets.
Its retrieval goes beyond similarity matching. RIVA preferentially activates fragments that have historically contributed to successful reasoning trajectories, an experience-driven bias that weights memories by demonstrated usefulness rather than surface relevance. At the system level, memory and reasoning interact bidirectionally and frequently: intermediate reasoning outputs are written back to memory in real time, and existing entries can be selectively overwritten or updated, so the memory space is continually refined rather than merely appended to. The cost side of this design is visible in the table above: recomposable reasoning demands complex memory operations, and the computational and storage overhead scales with them.
CEREBRA: Hierarchical and Governable Long-Horizon Memory 🍅
CEREBRA’s evolutionary path points the opposite way, toward the stability and governability of long-term knowledge. Its most distinctive feature is the spontaneous emergence of a hierarchical memory structure, in which different layers assume different semantic roles and information flow across layers is constrained by implicit gating mechanisms.
The important part is where this governance comes from. It is not handcrafted rules or a predefined schema; it is shaped gradually by task-level feedback signals. Through reward-driven iterative selection, the system evolves a layered structure that resembles how knowledge sediments and consolidates over time. The same feedback-driven consolidation that makes it stable also makes it slow to change: stronger structural inertia and a higher adaptation cost for newly introduced knowledge are the price of long-horizon reliability.
LIGHTWEIGHT: Minimalist Memory Under Resource Constraints 🍅
When resource constraints, explicit or implicit, enter the evolutionary objective, any disadvantage in computational and storage overhead gets amplified by selection. LIGHTWEIGHT is the memory form that emerges under this pressure. It deliberately abandons long-term knowledge accumulation and reasoning-trace reuse, restricting memory to short-lived, task-relevant caches. Writes happen only when the system judges the current information significantly valuable for subsequent decisions; retrieval is kept intentionally simple. The design minimizes the memory module’s interference with the overall reasoning process, and in exchange the system holds stable performance in resource-constrained environments, at the cost, as the table notes, of degenerating toward a short-term decision mechanism when the horizon lengthens.
MemEvolve: From Designed Memory Modules to Evolvable Structures 🍅
MemEvolve breaks the standing assumption that memory structures must be fixed by hand, promoting memory from a pre-designed component to an object that can be learned and evolved.
The current agent platform’s memory architecture is largely design-driven rather than evolution-driven. Drawing on MemEvolve, the platform’s memory structure could instead become task-characteristic-oriented, a structure that emerges from task demands rather than being predefined against an assumed workload.
Insight 1: Introduce the EvolveLab Four-Module Abstraction to Construct a Searchable Structural Space 🍅
Enhancing agent capability does not have to mean scaling the model or extending the context window. Structure itself can carry capability. When memory encoding schemes, storage strategies, retrieval paths, and inter-module connections are all subject to evolutionary search, performance gains stop depending solely on parameter expansion and start coming from structural reconfiguration and adaptation.
Historically, the agent platform’s capability growth has followed a linear expansion path:
model upgrades → RAG enhancement → GraphRAG → tool ecosystem expansion → context extension.
Each step on that path adds capacity; none of them adds structural adaptability, which is the dimension worth introducing next.
MemEvolve’s engineering foundation, EvolveLab, offers a practical reference for how: decouple standardized memory modules into independent, replaceable, evolvable units, turning memory architecture design into a combinatorial search problem over modular configurations.
For the agent platform, an EvolveLab-style modular decoupling has value even before any search runs: individual modules become replaceable and testable in isolation, without rewriting the whole system. That decoupling is the prerequisite for structural search and automated optimization later: it makes structural self-adaptation an incremental, engineerable evolution instead of a disruptive redesign.
Insight 2: From a Fixed Memory Paradigm to a Selectable and Evolvable Memory Space 🍅
The engineering lesson I take from MemEvolve is that committing to a single “correct” memory pipeline at design time is an unforced error. One can instead construct a searchable memory space and let different task distributions gradually shape their own suitable structures at runtime. MemEvolve’s outer loop shows what that takes in practice, a Diagnose-and-Design workflow: select candidate architectures via Pareto-based multi-objective ranking (task success rate, cost, latency); analyze failure cases and memory access patterns into a defect profile; then synthesize improved architectures from that diagnostic signal. The Pareto step matters because memory design has no single optimum: success rate, cost, and latency trade against each other, and ranking on one axis alone would collapse exactly the diversity the search needs.
For the agent platform, the raw material already exists: task logs carry failure traces, reasoning breakdowns, and memory access statistics that could be assembled into task-level defect profiles. Those signals could then drive targeted iteration across reasoning-oriented memory, hierarchical knowledge memory, and lightweight caching strategies. Compared to one-off manual design decisions, such a loop is more systematic, and better matched to what an agent platform, which sees many task distributions rather than one, actually needs.
This transition moves agent memory research toward self-organization and self-optimization, closer to the organically formed knowledge structures we expect intelligent systems to exhibit.
