remnic

Remnic is a lightweight, open‑source memory engine that lets you store and retrieve user context as plain markdown files on your local machine. It works next to your existing AI agents, Claude Code, Codex CLI, Cursor, ChatGPT via MCP, and many others, so that each agent shares the same single memory store.

When you launch an agent session, Remnic automatically loads relevant markdown files, extracts structured entries, and makes them available as a knowledge base for the next prompt. After the agent finishes its turn, Remnic records the new facts back into the same files, adding provenance fields such as source, timestamp, and scope. The storage format is simple: a markdown document coupled with a YAML front‑matter block. The files are managed locally, so you can view, edit, or archive them with any text editor, grep, or git. A daily snapshot of each file preserves version history, and diff tools allow you to revert to earlier states.

The system distinguishes between scopes, personal, work, client, project, repo, and temporary, so that context remains bounded. Each recall returns not just the text but also a confidence score, the source file, and the version that was used. Remnic's hybrid search blends BM25 keyword matching with vector similarity and optional graph‑based ranking; a “Recall X‑ray” view shows which tier generated each result. Additionally, provisions exist for correction: you can flag a memory as stale, which automatically removes it from future recall, or you can annotate a correction that propagates to the stored markdown.

For agents that run on MCP, Remnic provides a REST/MCP server that authenticates via OAuth 2.1. Native extensions exist for Claude Code, Codex CLI, Pi Coding Agent, Hermes, OpenClaw, and WeClone. A generic “MCP client” can pick up Remnic architecture without any extra changes. The command‑line tool @remnic/cli installs in under a minute and offers one‑line commands to install the daemon, start it, and interact with the memory store.

Key features

• Shared memory store in plain markdown + YAML across all MCP clients
• Scope‑based separation of personal, client, and project context
• Provenance metadata: source file, timestamp, confidence, and reason for recall
• Hybrid search with BM25, vector similarity, and optional graph ranking
• Local‑first design; files can be versioned, archived, or edited with standard tools
• SR&CD: structured recall, structured consolidation, date‑based supersession