Claude Code Memory System = CHEAT CODE
Jack Roberts presents a three-tier Claude memory system designed to give AI tools persistent context across all platforms and sessions. The system consists of short-term identity memory, mid-term project memory via structured folders and claude.md files, and long-term memory using either Obsidian or Pinecone for archiving conversations and expert knowledge. The goal is to eliminate the 'amnesia' problem where AI loses context between chats.
Summary
Jack Roberts, a tech entrepreneur and AI content creator, walks through what he calls a 'Claude memory system' — a structured three-level approach to giving AI assistants persistent, cross-platform memory. He frames the problem as AI amnesia: tools like Claude or ChatGPT lose context between sessions, leading to irrelevant or repetitive responses mid-conversation.
The first tier is short-term memory, which Roberts calls the 'operating manual.' This captures static personal identity information: name, role, goals, preferred tone, tools used, and non-negotiables. He recommends filling in Claude's native 'Instructions for Claude' field (under General settings) with up to 200 words of top-level context. In coding environments like VS Code or Antigravity, this maps to the global CLAUDE.md file.
The second tier is mid-term memory, structured around six to eight project folders — one per major area of life or business (e.g., agency, startup, health, community). Each folder contains a CLAUDE.md file that describes the project's goal, tech stack, prior decisions, and a memory map pointing to relevant references. Roberts recommends using Claude itself to generate these categories by asking it to organize your chat history into six to eight life/business areas. This layer is 'mutable' — it changes as projects evolve — and should be kept under 200 lines to avoid bloating every conversation.
The third tier is long-term memory, split into two components: a conversation archive and an expert knowledge base. For the archive, Roberts recommends ending every significant session with a 'wrap-up' skill that summarizes decisions, next actions, and metadata, then embeds that summary into Pinecone (a vector database) for semantic search later. For the knowledge base, he uses tools like FireCrawl (for deep web research) and NotebookLM (for curating AI-generated research notebooks on specific topics like business strategy), which can then be vectorized into Pinecone or stored in Obsidian.
Roberts contrasts the two long-term memory tools: Obsidian suits users who want to manually read, edit, and visually explore memories through graph views and backlinks, while Pinecone suits those who want scalable, semantically searchable storage accessible from anywhere. He personally uses Pinecone. The entire system is designed so that opening a new chat window in any platform still yields fully context-rich responses, because memory is imported at the prompt level rather than dependent on chat history.
Key Insights
- Roberts argues that the outcome of any AI conversation should never depend on chat history — a properly built memory system should allow you to open a brand new chat window and still receive fully context-rich responses, because memory is injected at the prompt level rather than inherited from prior sessions.
- Roberts frames memory not as a passive vault but as an active import, stating 'every prompt silently pulls from your stack — who you are, what you're shipping, what you decided last month' — meaning the response quality is structurally elevated before you even type a question.
- Roberts claims that asking Claude to analyze your own chat history and organize your life into six to eight categories is 'surprisingly accurate,' positioning the AI itself as the tool for bootstrapping the mid-term memory architecture.
- Roberts distinguishes Obsidian from Pinecone by use case: Obsidian is for users who want to manually read and edit memories in a visual graph format, while Pinecone is for semantic search across thousands of records and is better suited for storing large content like books and transcripts at scale.
- Roberts describes a 'wrap-up skill' that automatically summarizes an entire conversation — capturing decisions, next actions, and metadata — and embeds it into Pinecone at session end, enabling future retrieval of specific past discussions by date or topic filter.
Topics
Full transcript available for MurmurCast members
Sign Up to Access