zikra

Zikra is an MCP-native persistent memory server intended for use with Claude Code and other AI coding agents. The README frames it as a shared, governed memory layer for teams that run multiple agents, people, and projects. The server can run locally, binds to localhost:8000 by default, and uses Bearer token authentication. It can be added to Claude Code as an MCP server through a URL and Authorization header, and the project provides an interactive installer plus an update script. Key capabilities described include multi-agent support (Claude Code, Gemini CLI, Codex, or any agent that can POST), role-based access with owner, admin, developer, and viewer roles, and isolated project namespaces so one team can keep separate memory pools per project. Memory is structured rather than plain text: the command set covers saving memories and prompts, logging runs and errors, saving and listing requirements, promoting requirement types, and retrieving items by title or id. Search is described as hybrid semantic plus keyword when embeddings are enabled; without OPENAI_API_KEY the README says it falls back to keyword-only. Ranking considers age, access frequency, and confidence score. Storage can use SQLite for local or single-user setups, or PostgreSQL for team deployments. The PostgreSQL backend is configured through DB_BACKEND and DB_* environment variables, with an extra postgres install option. The repository description also mentions pgvector and n8n as part of the stack, and the documented webhook command endpoint is /webhook/zikra. Embeddings can use OpenAI or a compatible endpoint through OPENAI_API_BASE. Configuration options include token, database settings, bind address, port, default project, embedding model, decay days, and frequency weight. For Claude Code, setup prompts install Stop and PreCompact hooks plus a statusline bar, enabling automatic session logging. A built-in hygiene prompt is described for confidence decay, orphan detection, and stale cleanup. The project is MIT licensed.