JuliusBrussee/cavemem

What it solves

It provides cross-agent persistent memory for coding assistants, preventing AI agents from forgetting information across different sessions or when switching between different IDEs.

How it works

Cavemem uses hooks that fire at session boundaries to capture observations from the agent's activity. These observations are compressed using a deterministic "caveman grammar" (reducing prose tokens by approximately 75% while preserving code and paths exactly) and stored in a local SQLite database. Agents can then retrieve this history via a Model Context Protocol (MCP) server using hybrid search (combining SQLite FTS5 keyword search and a local vector index).

Who it’s for

Developers using AI coding assistants such as Claude Code, OpenCode, Codex, GitHub Copilot, Augment Code, Cursor, Gemini CLI, Antigravity, or IBM Bob who want a unified, local, and privacy-aware memory store for their agents.

Highlights

• Cross-IDE Support: Captures data from several IDEs and allows others to query that shared memory.
• Local-First: No network or cloud requirements by default; data is stored in a local SQLite DB.
• Compressed Storage: Uses a specialized grammar to save space while remaining human-readable upon expansion.
• Progressive Retrieval: MCP tools allow agents to search, view timelines, and fetch specific observations to manage context efficiently.
• Privacy Controls: Includes automatic redaction of secrets and support for path globs to exclude sensitive directories.
• Web Viewer: A local, token-protected UI for humans to browse captured sessions.