XNews

Claude memory, persistent memory projects and agent skills

Claude memory, persistent memory projects and agent skills

Comparative analyses and community projects detail Claude's retrieval-first memory, ChatGPT's injection summary approach, new 'skills' support, and Claude-Mem persistent memory tooling.

Key insights

Two dominant memory architectures: pre-injection vs on-demand retrieval

Reverse-engineering shows ChatGPT uses precomputed injection: session metadata, user long-term facts, recent conversation summaries, and current window content are injected each session. Claude defaults to an on-demand retrieval model with explicit tools (conversation_search, recent_chats) that the agent decides to call, plus XML-style long-term memory injection. The two approaches trade off immediacy and token cost (ChatGPT) vs precision and latency (Claude). (supporting posts: 17964)

ChatGPT adopts agent 'skills' mechanism inspired by Claude

ChatGPT reportedly added support for agent skills similar to Claude's model. A test using the Thinking model showed built-in skills for Excel, Word and PDF processing, enabling modular, composable skill usage that does not consume the context window and simplifies sharing and iteration of capabilities. (supporting posts: 17855)

Claude-Mem provides local persistent memory compression for Claude Code

Claude-Mem is an open-source persistent memory compression system for Claude Code that captures tool calls, compresses logs semantically, stores summaries in a local SQLite DB (backed by Chroma vectors), and injects relevant remembered items into new sessions. It aims to save token costs, enable cross-session continuity, and provide a local UI for memory inspection. (supporting posts: 17831)

Skill-based design improves composability and UI alignment for agents

Discussions note Anthropic's proposal to accumulate skills rather than rebuild agents repeatedly. Implementing skills yields stronger frontend results and allows domain-specific skill libraries (e.g., front-end design, coding), producing more consistent, environment-aware outputs. Skills can be assembled like LEGO blocks to create tailored agent behaviors. (supporting posts: 17747)

Source posts