Corpas Core + GBrain: five prompts for your first working AI memory system

The question behind Corpas Core is personal: how much more capable can I become by learning to think with machines?

As a scientist working across genomics and AI, I need continuity across papers, projects, conversations and decisions. I want an assistant that can recover why I chose a direction, recognise when the evidence has changed and help me finish work.

Corpas Core is the structure I am building around that ambition. GBrain supplies the knowledge layer. My current system also includes priorities, reusable procedures, approval rules and outcome records.

You can reproduce the basic pattern without reproducing my entire infrastructure. The prompts below aim to give you a small working foundation in one sitting. With the tools and access already configured, the starter can be assembled in minutes. Installing dependencies, connecting accounts and proving reliability take longer. This is a prompt-led starter, not a freshly benchmarked one-click installer or a claim that my complete system can be copied in ten minutes.

What you need

• A computer where an AI coding agent can read and write a dedicated local folder and use a terminal. This guide targets Codex or Claude Code.
• A working account for that agent, internet access and permission to install software on your machine.
• Git and Bun, or permission for your agent to install the missing prerequisite from its official source.
• GBrain from garrytan/gbrain. The unrelated npm package named gbrain is not the project used here.
• For semantic retrieval, a supported embedding-provider API key configured locally. GBrain also supports a keyword-only starting point. Model and embedding API usage can have separate costs from your coding-agent subscription.
• Three to ten non-sensitive notes that you are comfortable processing with your chosen services. Synthetic notes are enough for the first test.

Follow GBrain’s current installation guide for setup details. Locally stored files do not imply that every model call stays on your computer. Start with sample material and check the data flow before adding personal or institutional records.

Download the starter files and all five prompts. The kit contains templates and a fictional freshness test, not my private memory or production configuration.

Prompt 1: establish a working memory connection

Paste this into your coding agent in a new, dedicated workspace. Replace the bracketed fields first.

Set up a minimal personal knowledge workspace at [ABSOLUTE WORKSPACE PATH]. My agent is [CODEX OR CLAUDE CODE]. Use GBrain from https://github.com/garrytan/gbrain and read its current official installation and agent-connection documentation before choosing commands. Check for an existing installation and preserve any existing brain and agent configuration. If one exists, propose a separate source for this starter rather than reinitialising it. Otherwise use the simplest supported local setup. My retrieval choice is [KEYWORD ONLY / SEMANTIC WITH AN EXISTING LOCALLY CONFIGURED KEY]. Explain any missing prerequisite or paid service before using it. Never ask me to paste secrets into chat. Connect GBrain to this agent using its documented integration. Do not connect email, calendars or private folders, create a public repository, or enable recurring jobs. Verify that the agent can store a harmless sample note and retrieve it with its source identifier. Report the exact checks that passed and any remaining blocker. Do not call installation alone a working memory system.

The useful result is a successful retrieval through the agent you will actually use. A command that installed successfully is only an intermediate step.

Prompt 2: give the system priorities and boundaries

A memory store cannot decide which of your ambitions deserves attention. Supply that context explicitly.

In this workspace, create a small operating layer using plain Markdown files: HUMAN-OBJECTIVES.md, CURRENT-STATE.md, PROCEDURES.md and OUTCOMES.md. Use my answers below: objective [ONE SENTENCE]; current priorities [UP TO THREE]; constraints [TIME, DATA AND APPROVAL LIMITS]. Do not invent missing commitments or personal facts. Date the current-state file and identify who can change priorities. Add a scoped instruction file supported by my coding agent, preserving existing instructions, so substantive tasks read these records and retrieve relevant GBrain evidence first. Current priorities must come from CURRENT-STATE.md; older notes remain historical evidence. Refer to my objective file as human doctrine, distinct from any agent-personality SOUL.md that GBrain may use. Drafting and local analysis are allowed. External sending, publication and other consequential actions require my explicit authorisation. Record a task as complete only when its output and verification result exist. Unknown review time stays unknown.

These files provide a small, inspectable structure. They are a proposed starting design inspired by Corpas Core, not a description of every file in my production system.

Prompt 3: prove it can distinguish old information from current information

While preparing this post, GBrain returned an August architecture page that described an overdue planning cycle. A separate operational record generated on 4 September correctly showed the current cycle. Both existed. Only one answered the question about what was current.

Reproduce that problem deliberately with fictional material:

Use the starter kit’s fictional notes, or create two clearly labelled synthetic notes about the same project: an older plan with a Friday deadline and a newer approved state with a Monday deadline. Register only this sample material in the dedicated GBrain source, using the installed version’s documented commands. Preserve each note’s date, source and authority. Ask: “What is the current deadline, what changed, and which record supports the answer?” Retrieve and inspect the sources. The answer must identify Monday as current, explain the older Friday date and cite the newer approved record. It must not merge the dates or infer authority solely from search rank. Save the answer and the actual retrieved source identifiers to verification.md. If retrieval fails, report the failure without substituting your recollection of these instructions for retrieved evidence.

This is a basic acceptance check. Passing it does not prove that every future retrieval will be correct.

Prompt 4: turn memory into one useful piece of work

Choose a recurring task you already understand well enough to assess.

Help me perform [MEETING PREPARATION / RESEARCH-PAPER REVIEW / WEEKLY PROJECT REVIEW] using only [APPROVED NOTES OR DOCUMENTS]. First read my objectives and current state. Retrieve relevant evidence from GBrain and show the sources and dates. Identify contradictions or missing context before recommending action. Produce [SPECIFIC DELIVERABLE] in the workspace. Separate facts, inferences and proposed actions. Check the output against its sources and report unresolved uncertainty. Do not send or publish anything. Record the task, output path, sources, verification result and review status in OUTCOMES.md. Leave human review time blank until observed. If this task can be repeated, describe the procedure in PROCEDURES.md without claiming that it has already been validated across multiple cases.

The deliverable might be a one-page meeting brief with open commitments, or a paper review that separates reported results from your interpretation. It should be something you can inspect, not another dashboard about activity.

Prompt 5: retain a correction and test whether it helps

Review my correction to the last output: [CORRECTION]. Determine whether it belongs in a factual record, the current state or the reusable procedure. Propose the smallest change and apply it only within my existing authorisation. Preserve the original evidence and note what changed. Test the revised workflow on a second, different case that I supply. Compare the result with the correction and show whether the earlier mistake recurred. Record the outcome honestly, including failure. Do not mark the learning loop closed merely because a file was edited, and do not calculate time saved without a defensible baseline and observed review time.

This is where a memory system starts to become an execution system: a correction can alter later work, and you can inspect whether the change helped.

What this starter does not establish

My own current proof report does not demonstrate the overall “personal AGI” claim. My Cognitive Leverage Ratio is not yet measurable. Completed outputs and working procedures are useful evidence, but they do not justify saying I am ten times more productive.

Apply the same standard to your starter. Run several comparable tasks. Count preparation, review, correction and maintenance time. Check outcomes against sources. Keep failures visible. Expand the system only when the next integration solves a demonstrated problem.

For me, genomics is the proving ground. An assistant that remembers more must still preserve uncertainty and ask which populations an analysis represents. The broader ambition is human–AI cognitive augmentation; the practical beginning is one workflow that produces useful, checkable work.

What would your first workflow be, and what result would convince you that it had made you more capable?

Listen to the narrated article (9 minutes). Narrated with OpenAI Cedar, an AI-generated voice.