pax-beehive/paxm

What it solves

PAXM is a memory adaptor that prevents users from having to repeatedly explain project context, architecture decisions, and operational constraints to different AI coding agents. It provides a unified memory layer that allows context captured in one agent session (e.g., Codex, Claude Code) to be recalled in another, regardless of the agent or the underlying storage provider.

How it works

PAXM acts as a middleware between AI agents and memory providers. It supports two primary interaction paths:

• Active Path: Users can explicitly record or recall memories using a CLI, MCP (Model Context Protocol) server, or specific agent skills.
• Passive Path: Using lifecycle hooks, PAXM automatically recalls relevant context before an agent responds and captures completed conversation turns after they occur.

To ensure performance, passive writes are committed to a local durable queue and delivered to providers in the background, preventing provider latency from blocking the agent's response.

Who it’s for

It is designed for developers using AI coding agents (such as Cursor, Claude Code, Codex, and others) who want a persistent, cross-agent project memory that is independent of any single tool's built-in memory.

Highlights

• Cross-Agent Compatibility: Works across a wide array of agents including Codex, Claude Code, OpenCode, Pi, Cursor, TRAE, and Cline.
• Interchangeable Providers: Supports multiple backends including a zero-setup local SQLite provider, Zep, Mem0, MemOS, OpenViking, and custom JSON-RPC providers.
• Zero-Setup Local Memory: The built-in SQLite provider uses FTS5 and BM25 retrieval without requiring API keys, embeddings, or external LLM calls.
• Reliability: Features background retries for provider delivery, timeouts to prevent agent blocking, and a durable local queue for writes.