Show HN: CogniKernel; AI Coding with cross session and cross platform memory 展示 HN：CogniKernel；具有跨会话和跨平台内存的 AI 编码

TL;DR

• CogniKernel provides persistent, structured project memory for AI coding agents like Claude Code and Codex by extracting decisions, constraints, and abandoned approaches from coding sessions.
• It uses an event-sourced store with lexical-primary retrieval and write-time consolidation to inject compact context blocks into future sessions, preventing redundant re-decisions.
• The system avoids LLMs in the loop by using two small fine-tuned encoder models (~130 MB ONNX each) running locally on CPU for salience classification and supersession detection.
• Memory is keyed by project path, enabling cross-tool consistency between Claude Code and Codex within the same directory.
• The architecture includes a fail-open design ensuring session continuity even if optional model components are missing, degrading gracefully to lexical fallbacks.

Why It Matters

This addresses a critical pain point in AI-assisted coding: context fragmentation across sessions where agents repeatedly rediscover or contradict prior decisions. By implementing local, deterministic memory extraction without cloud dependencies or per-session token costs, CogniKernel offers a privacy-preserving, cost-efficient alternative to vector-database-based memory systems that rely on generative summarization. Its approach demonstrates how lightweight specialized models can outperform heavy LLM loops for specific classification tasks while maintaining reliability through fail-open engineering.

Technical Details

• Event-Sourced Architecture: Maintains an immutable log of typed memory events (DECISION, CONSTRAINT_HARD/SOFT, APPROACH_ABANDONED) rather than mutable state, enabling auditability and replayability of project evolution.
• Dual-Model Pipeline: Uses salience_v2 (SetFit classifier with bge-small backbone) to classify sentence relevance/type during session capture, and supersession_xenc (cross-encoder) to detect when new information overrides existing entries via latest-wins consolidation.
• Lexical-Primary Retrieval: Combines keyword-based search with dense embeddings from the encoder models, prioritizing exact matches before falling back to semantic similarity to ensure predictable recall behavior.
• AST Skeleton Injection: Projects abstract syntax tree structures into agent context windows instead of raw file contents, reducing read operations by 60-80% according to internal benchmarks while preserving structural understanding.
• MCP Tool Integration: Exposes four targeted commands (recall, find_related, skeleton, get_session_state) for programmatic memory access within coding environments, supporting both interactive and scripted workflows.

Industry Insight

The tool's success highlights a strategic shift toward "local-first" AI memory solutions as enterprises prioritize data sovereignty and operational resilience against API failures or pricing volatility. Developers should consider integrating similar event-sourced pattern recognition into their agent frameworks rather than relying solely on embedding databases, particularly for domains requiring strict compliance (e.g., healthcare or finance code generation). Future iterations may expand beyond coding contexts to technical documentation or research note-taking systems where maintaining decision provenance becomes equally valuable.

TL;DR

• CogniKernel 是一个为 Claude Code 和 Codex 设计的持久化、结构化项目记忆系统，通过事件溯源存储提取编码会话中的决策与约束。
• 它采用本地运行的轻量级编码器模型（~130 MB ONNX）进行分类而非生成，实现毫秒级处理且无需 API 调用或云端传输。
• 核心特性包括跨工具记忆共享、行动时护栏机制、AST 骨架读效率优化以及 fail-open 可靠性设计。
• 安装支持 PyPI 与 uv 工具链，初始化后可自动捕获并注入上下文，减少重复决策与文件读取开销。
• 模型基于真实项目数据验证，保留人类标注一致性上限，确保稀有类别的持续成长与鲁棒性。

为什么值得看

该方案重新定义了 AI 编程助手的“记忆”范式——不依赖 LLM 摘要或向量数据库，而是通过确定性分类 pipeline 实现高效、隐私安全的本地知识留存。对开发者而言，这意味着更低的成本、更快的响应速度以及更强的上下文连贯性，尤其适合长期迭代型项目。同时其 fail-open 架构保障了系统在模型缺失时的可用性，具备工程落地的高可靠性。

技术解析

CogniKernel 的核心在于将记忆抽取视为分类问题而非生成任务：使用两个小型 fine-tuned encoder 模型（salience_v2 与 supersession_xenc）分别判断语句是否值得记忆及是否覆盖旧记录。salience_v2 基于 bge-small backbone，以 SetFit 架构训练，输出 DECISION / CONSTRAINT_HARD/SOFT / APPROACH_ABANDONED_DO_NOT_RETRY 等标签；supersession_xenc 则作为 cross-encoder 检测语义冲突。所有推理在本地 CPU 完成，延迟低至毫秒级，避免 token 费用与数据外泄。存储层采用事件溯源日志结构，按 project path 键值组织，支持多个 IDE 工具共享同一记忆库。此外，系统提供 MCP 接口如 recall、find_related、skeleton 等供显式调用，并通过 AST 骨架注入提升后续会话的阅读效率。

行业启示

随着 Agent 编程场景深化，“记忆成为第一优先级基础设施”的趋势愈发明显。CogniKernel 证明无需大模型也能构建高可用记忆系统，推动行业向“轻量化 + 本地化 + 可解释”方向演进。未来应关注此类模块化组件如何嵌入主流开发环境（如 VS Code、JetBrains），并形成标准化协议促进跨平台记忆互通。同时，其 fail-open 设计理念也为关键业务系统的容错机制提供了新参考——即使智能模块降级，核心功能仍可稳健运行。