CogniCore
Provide persistent memory, cross-session handoff, reflection, and safety guardrails for AI coding agents without requiring external vector databases.
CogniCore is a lightweight, local-first library and MCP server designed to give AI agents persistent memory, reflection capabilities, and safety guardrails across sessions. It enables seamless context and task handoffs between different coding agents—such as Claude Code, Codex, and Cursor—without re-explaining decisions, tickets, or rejected approaches. Operating entirely in-process using SQLite and multi-hop BM25 retrieval, CogniCore eliminates the need for external vector databases or embedding servers while ensuring strict privacy by redacting secrets and ignoring raw tool outputs.
Características Principales
01Reflection engine that records past failures and extracts lessons to avoid repeating dead ends
02Deterministic replay, task branching, and cryptographically signed memory export bundles
03In-process memory and multi-hop BM25 retrieval without requiring external vector databases or embedding servers
0484 GitHub stars
05Seamless session handoff and context sharing across AI agents like Claude Code, Codex, and Cursor
06NexusShield guardrails to intercept and block prompt injection and jailbreak attempts
Casos de Uso
01Resuming complex software debugging sessions in different coding agents without losing past context or attempted fixes
02Preventing prompt injections and jailbreaks in autonomous coding workflows using built-in agent safety shields
03Transferring verified, cryptographically signed memory states between disparate agent frameworks like mem0