基于人类认知的 LLM 智能体长期记忆架构 A Human-Cognitive-Inspired Long-Term Memory Architecture for LLM Agents
Memory Life 是一种面向基于 LLM 的智能体的长期记忆系统，它建模人类认知记忆机制——不是作为被动存储，而是作为主动认知状态，塑造行为、形成人格并自我管理其生命周期。
Memory Life is a long-term memory system for LLM-based agents that models human cognitive memory mechanisms — not as passive storage, but as active cognitive state that shapes behavior, forms personality, and self-manages its own lifecycle.
核心能力 / Key capabilities：
- 以价值为中心的保留 / Value-centric retention：只有具有前瞻性影响的信息被保留；闲聊和噪音被自动过滤。/ Only information with forward-looking impact is retained; chitchat and noise are automatically filtered.
- 三层意识模型 / Three-layer consciousness：显式召回 → 自动化行为默认 → 蒸馏人格，实现跨会话的渐进式人格形成。/ Explicit recall → automated behavioral defaults → distilled persona, enabling progressive personality formation across sessions.
- 分代记忆垃圾回收 / Generational memory GC：受 JVM 启发的记忆生命周期管理，带有从暂存区到永久类别的经验晋升。/ JVM-inspired memory lifecycle management with empirical promotion from tentative to permanent categories.
- 情境化知识 / Contextualized knowledge：冲突通过范围标注解决，而非破坏性覆盖。/ Conflicts resolved by scope annotation, not destructive overwrite.
详见 ARCHITECTURE.md 了解完整的技术设计，包括新颖的价值评估模型、意识层、GC 机制和运行工作流。
See ARCHITECTURE.md for the complete technical design, including the novel Value assessment model, consciousness layers, GC mechanism, and operational workflow.
本仓库包含 Memory Life 架构的第一版实现（v0.1），作为参考和起点提供。
This repository contains the first implementation (v0.1) of the Memory Life architecture, provided as reference and starting point.
GPL-3.0。详见 LICENSE。
GPL-3.0. See LICENSE.
详见 CONTRIBUTING.md。
See CONTRIBUTING.md.