为 AI 代理提供持久内存。开源、不依赖大型语言模型，支持任何 MCP 客户端。

星级历史

它的作用

大多数AI代理在会话结束时会忘记所有内容。cortex-engine解决了这个问题——它为代理提供了一个跨会话、模型和运行时的持久内存层。

• 语义记忆——存储和检索观察、信念、问题和假设作为互联节点
• 信念追踪 — 代理人持有当新证据反驳时更新立场
• 两阶段梦境巩固 — NREM压缩（聚类、精炼、创建）+ 快速眼动睡眠整合（连接、评分、抽象）——以生物睡眠阶段为模型
• 目标导向认知 — goal_set创造期望的未来状态，产生前瞻性预测误差，从而倾向于巩固和探索，朝向重要内容
• 神经科学基础检索 — GNN邻域聚合、查询条件扩散激活、多锚千脑投票、认知采集
• 信息几何——局部自适应聚类阈值，尊重嵌入空间曲率，模式一致性评分

@fozikio/cortex-engine

Persistent memory for AI agents. Open source, LLM-agnostic, works with any MCP client.

Star History

What It Does

Most AI agents forget everything when the session ends. cortex-engine fixes that — it gives agents a persistent memory layer that survives across sessions, models, and runtimes.

• Semantic memory — store and retrieve observations, beliefs, questions, and hypotheses as interconnected nodes
• Belief tracking — agents hold positions that update when new evidence contradicts them
• Two-phase dream consolidation — NREM compression (cluster, refine, create) + REM integration (connect, score, abstract) — modeled on biological sleep stages
• Goal-directed cognition — goal_set creates desired future states that generate forward prediction error, biasing consolidation and exploration toward what matters
• Neuroscience-grounded retrieval — GNN neighborhood aggregation, query-conditioned spreading activation, multi-anchor Thousand Brains voting, epistemic foraging
• Information geometry — locally-adaptive clustering thresholds that respect embedding space curvature, schema congruence scoring
• Graph health metrics — Fiedler value (algebraic connectivity) measures knowledge integration; PE saturation detection prevents identity model ossification
• Spaced repetition (FSRS) — interval-aware scheduling with consolidation-state-dependent decay profiles