docs: clarify cognitive memory strategy

README.md

@@ -29,6 +29,27 @@ DNA Memory 只保存有界来源指针、经过审查的提案，以及验证后
 它不是另一个无限增长的聊天数据库。原始会话只是形成认知的材料；只有经过筛选、
 验证并值得在未来任务中复用的结论，才会进入长期记忆真源。
 
+## 核心策略：先处理记忆，再存储记忆
+
+DNA Memory 的核心不是把更多对话塞进数据库，而是给 Agent 一套接近人类记忆
+形成过程的处理策略。Agent 把对话当作原始感知输入，只有经过策略闸门的内容，
+才可能成为可复用的长期认知：
+
+1. **感知边界**：只读取有界来源和必要的生命周期元数据，不把完整 transcript
+   当作长期记忆。
+2. **信号筛选**：过滤操作话术、临时状态、凭证和无法复用的噪声。
+3. **加权判断**：结合重要性、置信度、来源和后续使用反馈，判断一条信息是否值得保留。
+4. **认知分型**：把信号分为偏好、事实、洞察、决策、项目状态、开放事项、工作流和错误教训。
+5. **验证结晶**：只有安全、短小、可追溯且经过验证的结论，才写入 Markdown 真源。
+6. **召回与反馈**：在新任务开始前按关键词召回；只有真正影响任务的结果才反馈
+   `useful` 或 `misleading`。
+7. **演化与遗忘**：重复认知被精确去重，失效认知通过显式 `supersedes` 替代，
+   原始来源指针和过期候选按保留策略清理。
+
+这是一种受仿生学启发的工程策略，不声称复制人脑。它的目标是让 Agent 具备
+“筛选—加权—抽象—验证—召回—修正”的记忆行为，而不是拥有一个不断膨胀的聊天
+存档库。DNA Memory 的质量应看高质量认知是否被正确复用，而不是看保存了多少文本。
+
 ```text
 大量原始对话
   -> 有界采集与信号提取

README_EN.md

@@ -34,6 +34,35 @@ It is not another indefinitely growing chat database. Native conversations are
 source material; only conclusions that have been filtered, verified, and are
 worth reusing in future work become durable memory.
 
+## Core Strategy: Process Memory Before Storing It
+
+DNA Memory is not designed to push more conversation into a database. Its core
+is a policy loop that gives an agent a human-memory-inspired way to process
+experience. Conversations are treated as raw perceptual input; only signals
+that pass the policy gates can become reusable long-term cognition:
+
+1. **Bound perception**: read bounded sources and necessary lifecycle metadata,
+   not complete transcripts as durable memory.
+2. **Filter signals**: remove operational talk, transient status, credentials,
+   and conclusions that cannot be reused.
+3. **Weight evidence**: combine importance, confidence, provenance, and later
+   usage feedback to decide whether a signal is worth keeping.
+4. **Classify cognition**: map signals to preferences, facts, insights,
+   decisions, project state, open loops, workflows, and error lessons.
+5. **Crystallize after verification**: write only safe, short, traceable,
+   verified conclusions to the Markdown truth store.
+6. **Recall and learn from feedback**: recall before context-dependent tasks;
+   record `useful` or `misleading` only when the result actually affected work.
+7. **Evolve and forget**: exact duplicates are collapsed, invalidated cognition
+   is replaced through explicit `supersedes`, and stale pointers are retained
+   only according to bounded retention rules.
+
+This is an engineering strategy inspired by cognitive processes, not a claim
+to reproduce the human brain. The goal is a repeatable
+“filter-weight-abstract-verify-recall-correct” loop, rather than an ever-growing
+chat archive. DNA Memory should be evaluated by whether high-quality cognition
+is reused correctly, not by how much text it stores.
+
 ```text
 raw conversations
   -> bounded collection and signal extraction