Build a Portable Claude Memory and Automation System
Create a local knowledge folder Claude can read and update, then add connectors, reusable skills, recurring tasks, schedules, and agent workflows with human approval points.
Download the complete Claude workflow guide
Build a Portable Claude Memory and Automation System
Set up connectors, portable memory, reusable skills, recurring tasks, and agent workflows in one practical system.
What you will achieve
You will create a local AI brain that Claude can reference and update, then use it to run repeatable tasks with skills, loops, schedules, and orchestrated agents.
Who this is for
Claude users who want more personalized assistance, repeatable automation, or scalable business workflows.
Difficulty: Intermediate
Short tutorial
Set up connectors, portable memory, reusable skills, recurring tasks, and agent workflows in one practical system.
Guide: Build a Context Aware Claude Workflow
Time: 45 to 60 minutes
What You Will Build
This tutorial turns Claude from a blank chat window into a context aware working system. You will create a portable memory folder that you own, connect relevant data sources, teach Claude how you perform recurring tasks, and configure work that can run repeatedly.
Complete the steps in order. Start locally so you can inspect every file and understand how the system works. Add cloud storage or more advanced orchestration only after the local version is reliable.
1. Choose the Right Claude Surface
Use the interface that matches the work you are doing.
| Surface | Best for | Context available |
|---|---|---|
| Claude Chat | Questions, brainstorming, and document review | The current conversation, project context, and connected sources |
| Claude Cowork | Working with local folders, documents, and editable files | The folder you explicitly make available |
| Claude Code | Larger builds, application development, and coding workflows | Project files and development context |
For this setup, start in Claude Cowork. It retains a familiar conversational interface while allowing Claude to read and edit your local memory files.
Use Chat when you are on the go or only need a quick answer. Move to Claude Code when the task becomes a substantial software build.
2. Add Relevant Connectors
Connectors let Claude access specialized information or applications instead of relying only on the model's existing knowledge.
Open Claude settings, find Connectors, and search the available connector library. Add only the sources that support work you perform regularly.
Examples mentioned in this workflow include:
- Google Drive for company documents
- Gmail for email related work
- Slack for team information
- Google Calendar for scheduled commitments
- Notion for tasks and workspace content
- FMP or Quiver for financial and market data
- Media generation services when content production is part of the workflow
A connector is most useful when it supplies current or specialized data. For example, a financial data connector can provide earnings information that Claude can analyze against the portfolio context in your memory system.
Connector setup checklist
- Open Claude settings.
- Select Connectors.
- Search for the application or data source.
- Complete its authorization process.
- Start a test conversation that explicitly asks Claude to use the connected source.
- Confirm that the response references the correct source and current information.
Do not add connectors merely because they are available. Each one should support a defined workflow.
3. Create a Portable Local Memory System
The most important improvement is giving Claude accurate context that you control. Store that context in ordinary Markdown files so it can be reviewed, edited, backed up, and moved to another model later.
Create this folder on your computer:
AI-Brain/
context.md
instructions.md
memory.md
domains/
business/
personal/
skills/The three core files have different purposes:
| File | Purpose |
|---|---|
context.md | Stable facts about you, your work, objectives, constraints, and preferences |
instructions.md | Rules describing how Claude should respond, retrieve context, and update files |
memory.md | Important information that changes over time, such as new clients, strategic decisions, or active priorities |
The domains folder holds detailed information that Claude should load only when it is relevant. This prevents every request from carrying your entire business history.
Generate context.md through an interview
Start a voice or text conversation in Claude Cowork and paste this prompt:
Help me create an accurate context.md file for a portable AI memory system. Interview me about: 1. Who I am and what I do. 2. My business or professional responsibilities. 3. My current objectives and priorities. 4. My constraints, resources, and important financial assumptions. 5. How I make decisions. 6. My communication and output preferences. 7. The projects, products, clients, and audiences that matter most. Ask one focused question at a time. Follow up whenever an answer is vague or incomplete. Do not invent facts or fill gaps with assumptions. When the interview is complete, draft context.md in concise Markdown. Organize it with clear headings and bullet points. Flag any statement that still needs verification.
After Claude drafts the file, review it line by line. This manual check is essential because future recommendations may depend on these facts.
Verify especially:
- Financial figures
- Dates and deadlines
- Business names
- Client details
- Strategic priorities
- Personal constraints
- Claims transcribed from voice input
A transcription error that turns $100,000 into $1,000,000 could distort every later financial recommendation.
Create instructions.md
Use the following as a starting template:
# Claude Working Instructions ## Response Style * Be concise, analytical, and strategic. * Do not ramble or repeat the request. * Do not confirm my bias merely to be agreeable. * Identify weak assumptions, risks, and missing information. * Prefer specific recommendations over generic advice. ## Context Use * Read context.md for stable information about me and my work. * Read memory.md when the request may depend on recent decisions, priorities, or events. * Read files inside domains only when the request relates to that domain. * Do not load every domain folder for unrelated questions. ## Memory Updates * Update memory.md whenever I provide a substantial new fact, decision, objective, constraint, or change in direction. * Do not add casual conversation, duplicated information, or temporary details with no future value. * Notify me every time you change one of my files. * In the notification, name the file and summarize what changed. * Never silently replace an existing fact when the new information appears contradictory. Ask me to resolve the conflict. ## Accuracy * Do not invent missing context. * Distinguish confirmed facts from assumptions. * Ask for clarification when an uncertain fact could materially affect the answer.
Initialize memory.md
# Working Memory ## Current Priorities * Add current priorities here. ## Recent Decisions * Add material decisions here. ## Active Projects * Add projects whose status affects future work. ## Important Changes * Add new clients, offers, responsibilities, constraints, or strategic changes here. ## Open Questions * Add unresolved issues that may require a later decision.
Open the AI-Brain folder in Claude Cowork so Claude can read and edit these files.
Keep detailed context conditional
Store large bodies of information inside domain folders. For example:
AI-Brain/
domains/
business/
overview.md
products.md
funnel.md
clients.md
strategy.mdThe instruction to read these files only when relevant keeps the core memory smaller, reduces unnecessary context use, and lowers the chance of confusing unrelated information.
Maintain the memory
When your situation changes, tell Claude directly:
I have landed a new client called Northstar. The engagement begins next month and changes my available production capacity. Treat this as a material business update. Update the appropriate memory file, tell me exactly what you changed, and flag any existing information that now conflicts with this update.
Periodically compact the memory with this prompt:
Review memory.md for bloat, duplication, stale details, and contradictions. Propose a shorter version that preserves every fact still needed for future decisions. Do not edit the file yet. First show me: 1. What you recommend removing. 2. What you recommend merging. 3. What appears contradictory or uncertain. 4. The proposed revised file. Wait for my approval before applying the changes.
Back up the folder regularly. A cloud copy can protect the system if your computer or local files fail. If you later need shared access across applications, you can consider a cloud memory system, but the local folder is the clearest place to begin.
4. Turn Repeated Work into Skills
Memory tells Claude about you. A skill tells Claude how you perform a specific task.
Good candidates include:
- Video scripting
- Competitor research
- Copywriting
- Lead generation
- Content ideation
- Property research
- Opportunity scanning
- Repetitive data entry
Start with a task you already perform often and understand well.
Create a skill through an interview
Where the skill creator command is available, begin with:
/skill-creator
Then provide this brief:
Create a reusable skill for my competitor research process. Interview me to learn: 1. The objective and required final output. 2. The sources I inspect. 3. The exact sequence I follow. 4. The criteria I use to judge strong and weak examples. 5. The calculations or comparisons I make. 6. The mistakes that should cause rejection. 7. My preferred output format. 8. The situations that require human review. Ask for sample files or previous outputs when they would improve the skill. Do not finalize the skill until you can describe my process back to me accurately. The finished skill must include: * Trigger conditions * Required inputs * Ordered procedure * Decision criteria * Output format * Quality checks * Escalation conditions
Teach the task by demonstration
If your interface includes screen recording, record yourself completing one full example while narrating your reasoning.
Explain:
- What you are doing at each step.
- Why you choose one option over another.
- What signals indicate quality.
- What causes you to reject an item.
- What exceptions alter the normal process.
- What the final output should look like.
This is especially useful when a spreadsheet or document alone would not reveal your judgment criteria.
Use a manual skill specification when needed
# Skill: Competitor Research ## Objective Identify proven content opportunities without copying competitors. ## Inputs * Target topic or search term * Competitor list, if available * Time period to inspect * Desired number of recommendations ## Procedure 1. Search the specified topic. 2. Identify relevant competing channels or sources. 3. Review the strongest recent examples. 4. Compare performance, topic, framing, title, and format. 5. Separate isolated outliers from repeatable patterns. 6. Generate original ideas based on the strongest patterns. 7. Rank the ideas by confidence. ## Rejection Criteria * The idea depends on copying another creator. * The evidence is weak or based on one unexplained outlier. * The topic is not relevant to the target audience. * The available evidence does not support the claimed opportunity. ## Output Return a ranked table containing: * Idea * Supporting evidence * Recommended angle * Confidence * Risks or uncertainties ## Human Review Request review when evidence conflicts, source data is incomplete, or a recommendation could materially affect brand positioning.
After each use, give specific feedback and ask Claude to update the skill:
The process was correct, but the recommendations were too broad. Update this skill so future ideas must identify a specific audience problem, a distinct angle, and the evidence supporting that angle. Show me the proposed skill changes before applying them.
5. Add Goals, Loops, and Schedules
A skill still needs to be invoked. Loops and schedules move suitable tasks toward recurring execution.
Use a goal to define the desired result, then a loop to define what should repeat and how often.
Recurring content research example
/goal Maintain a ranked list of the five strongest video ideas based on topics currently performing well on YouTube. /loop Every 24 hours: 1. Review top ranked videos for these search terms: Claude, AI workflows, AI automation. 2. Use my competitor research skill and its exact decision criteria. 3. Identify repeatable patterns rather than copying individual videos. 4. Update the connected spreadsheet. 5. Rank ideas from highest to lowest confidence. 6. Include evidence, proposed angle, confidence, and risks. 7. Notify me when the list changes materially.
Habit dashboard example
/goal Improve my productivity using a personal habit tracking dashboard. /loop After each new set of daily entries: 1. Review adherence, missed habits, and recurring patterns. 2. Identify one change most likely to improve consistency. 3. Adjust the dashboard only when the evidence supports the change. 4. Record what changed and why. 5. Do not remove historical data.
Scheduled market report example
This workflow assumes the appropriate financial connector and portfolio context are available.
Goal: Give me a concise morning report containing only market information relevant to my portfolio. At the scheduled time: 1. Use the connected financial data source to review overnight market movements and relevant company updates. 2. Read my portfolio context from the appropriate memory file. 3. Identify material earnings information, market movements, and risks. 4. Separate confirmed facts from interpretation. 5. Explain why each included item matters to my holdings. 6. Send the finished report to the configured destination. Format: * Portfolio impact summary * Material company updates * Relevant market movements * Risks to monitor * Decisions that may require human review
Use loops for tasks that repeat at shorter intervals or react to new entries. Use schedules for work that should happen at a specific daily or weekly time. A task does not need both.
6. Design an Agent Workflow for Complex Processes
For a larger process, break the work into a graph of steps. Assign each step to a suitable model or agent, and show where a person must approve the result.
A stronger model can act as the orchestrator, while faster or less costly models handle simpler subtasks. This can reduce token use without assigning difficult judgment work to an unsuitable model.
Workflow mapping template
| Step | Input | Assigned worker | Output | Human approval |
|---|---|---|---|---|
| Ideation | Research and audience context | Research agent | Ranked ideas | Yes |
| Outline | Approved idea | Writing agent | Structured outline | Optional |
| Script | Approved outline and voice skill | Writing agent | Draft script | Yes |
| Quality review | Draft and brand rules | Review agent | Issues and revisions | No |
| Final approval | Revised script | Human | Approved script | Yes |
| Publishing preparation | Approved assets | Operations agent | Upload package | Yes |
Orchestrator prompt
Act as the orchestrator for the workflow defined below. Objective: Produce a publication ready video package from research through final preparation. Rules: 1. Break the objective into the documented workflow steps. 2. Assign each subtask to the least costly available model that can complete it reliably. 3. Use stronger reasoning for strategy, ambiguous decisions, and final quality review. 4. Use the relevant memory files and skills only when needed. 5. Preserve the output of every completed step for the next worker. 6. Stop at every designated human approval point. 7. If a subtask fails its quality criteria, revise or rerun it before proceeding. 8. Report which worker completed each step and what files changed. Workflow: [Paste the workflow table or process specification here] Success criteria: [Paste the measurable requirements for the final deliverable here]
Once the workflow performs reliably, combine it with an appropriate loop or schedule. Keep human approval points anywhere the process affects strategy, finances, public publishing, or another consequential decision.
Final Configuration Check
Before relying on the system, confirm that:
context.mdcontains verified facts rather than assumptions.instructions.mdrequires change notifications.memory.mdcontains only material, reusable information.- Detailed domain files are loaded only when relevant.
- Claude can access the selected local folder.
- Each connector has passed a source specific test.
- Each skill has defined inputs, steps, rejection criteria, and output format.
- Loops specify an interval and a clear update destination.
- Scheduled tasks define their source, format, and delivery target.
- Complex workflows include explicit human approval points.
- The complete memory folder has a current backup.
