Claude-Mem is an open-source 📝Claude Code plugin that provides persistent session memory by deploying a dedicated observer AI that watches coding sessions and captures structured observations in real time. Built by thedotmack, it installs via the Claude Code plugin marketplace and stores every decision, bug fix, architectural choice, and discovery as a searchable, categorized log — making prior session context retrievable without manual documentation.
The system uses a two-AI architecture: one AI codes, another watches and takes notes. Observations are auto-categorized by type (decision, bugfix, feature, discovery) and scoped by file path or semantic concept. Queries like type:decision file:auth.ts or concept-level lookups like "what do we know about token refresh?" both resolve against the same observation store.
Progressive disclosure keeps token usage efficient: sessions load a lightweight index (titles, types, timestamps, ~40 tokens each) and the LLM fetches full observations only when depth is needed. Each full observation (~850 tokens) includes before/after context — what preceded the observation and what followed — so causality is preserved, not just snapshots.
Installation:
/plugin marketplace add thedotmack/claude-mem && /plugin install claude-memThe team is working toward RAD (Retrieval Augmented Development) — an open standard for AI agent working memory analogous to what RAG did for external knowledge retrieval. Hook-based architecture, intelligent compression, and temporal awareness are the stated design targets.
- GitHub: github.com/thedotmack/claude-mem
- Docs: docs.claude-mem.ai
This is directly adjacent to what MythOS's MCP get_context call does at session start — but scoped to Claude Code specifically, and framed as an automated observer rather than a structured knowledge base. The "one AI takes notes on another AI" framing is clever and the categorization schema (decision / bugfix / feature / discovery) is a reasonable default taxonomy for coding sessions.
The before/after causality chain is the most genuinely differentiated feature — most memory systems capture what happened, not why it happened in sequence. That's where debugging value lives.
The RAD concept is worth tracking. If it gains adoption as an open standard, it could either be a layer MythOS integrates or a competing paradigm. The framing is smart — analogizing to RAG gives it an obvious narrative hook for the developer audience.
The core gap is that Claude-Mem is session-scoped and domain-specific (coding). MythOS is identity-scoped and domain-general. They're solving adjacent but different problems: Claude-Mem is episodic coding memory; MythOS is semantic life/work infrastructure. They could coexist — Claude-Mem observations could potentially be pulled into MythOS as structured memos.