Memory That Travels With You
Your best operations lead leaves in March. Her replacement starts in June and asks the company AI how the team handles quarterly reviews. The AI answers with generic advice. The useful memory lived in her old chat history, which was deleted along with her seat.
Meanwhile your company is switching AI tools. IT announces the migration. Someone asks what happens to two years of saved memories, preferences, and instructions. The honest answer is that nobody knows, because nobody owns them.
That is the cost of memory that belongs to the product instead of the business.
The 60 second version
- Treat the AI's built in memory as a convenience, not an archive. Assume it will be wiped.
- Keep the knowledge your company needs in plain files you own: standards, decisions, terminology, lessons.
- Screen material for proprietary data before any upload. The projector test decides.
- Update through an approval loop: chat proposes, you approve, the file records, the changelog explains.
- Carry the files to every new tool. Memory travels. Accounts do not.
Account memory is rented. Files are owned.
The contrast pair from Part 1, fragile versus portable, applies to company knowledge too.
Built in memory is fast and effortless. It remembers you prefer short emails and knows the team's product names. For low stakes convenience it is useful. Use it for that.
It cannot promise to survive a tool switch, an IT policy change, a license reassignment, or an employee departure. Its controls are product specific: what one tool calls memory, another calls instructions, and a third calls context. A rule saved in one product may not transfer to the next, and the next model may reinterpret it anyway.
Memory is product specific. Files are yours.
The operating rule
Keep the memory that matters in files your company controls, in an approved shared location. When the tool changes, the files move with you. When a person leaves, the knowledge stays.
Seed the system from work you already have
You do not build organizational memory by typing lessons into a chat. You build it by collecting material that already holds the knowledge: proposals, briefs, decision memos, postmortems, the explanations your best people give new hires. The method below mirrors the intake process in Part 1: Teach AI Your Voice Once, Use It Forever, applied to company knowledge instead of writing voice.
Step 1: Create a temporary intake folder
Put candidate files in incoming/ first. Nothing reaches the permanent library without review.
Company_Knowledge/
README.md
STANDARDS.md
DECISIONS.md
TERMINOLOGY.md
LESSONS.md
CHANGELOG.md
incoming/
library/Step 2: Screen for proprietary data before uploading
This is the step people skip. Review every file for material that must never reach an AI system:
- Customer records and contact details
- Credentials, access keys, and private links
- Internal financial results and deal terms
- Unreleased product or strategy plans
- Regulated personal or health information
- Client names hiding behind weak pseudonyms
A weak pseudonym is "large regional manufacturer" when everyone in the room knows there was only one. Sanitization must remove the ability to work backward to the source.
The projector test: if the content would embarrass or endanger you shown on a projector to strangers, it does not belong in the upload. Replace real facts with fictional equivalents that preserve the lesson: a three step explanation of why an approval process delayed revenue becomes a fictional company, invented numbers, and the same structure. The system is learning how the company thinks, not what the company owns.
Also check the platform before uploading anything:
- Is the tool approved for this material?
- Who can access the shared folder or project?
- How long are uploaded files retained?
- Can the files be removed later?
If you cannot answer, use invented material.
Step 3: Ask for an inventory, not conclusions
Review the attached files as candidate knowledge for a company memory library.
Do not draft anything yet. For each file, identify:
1. Its intended audience and business purpose.
2. Repeatable traits: standards, decision frameworks, terminology, process lessons that could apply to future work.
3. Traits specific to that assignment: names, dates, deal specifics, context that should not be generalized.
4. Anything outdated, contradicted by another file, or superseded by a later decision.
5. Any content that may be confidential, proprietary, personal, or identifying, even if disguised.
Cite the filename and a short excerpt for every finding.The AI shows its work. You decide what stays.
Step 4: Promote ten strong, not fifty mixed
Move the best sanitized files from incoming/ to library/. For each one, add a short note: why it belongs, which lessons the AI should reuse, which details are specific to that case, anything you would update today.
Ten strong beats fifty mixed. A library of obsolete process docs teaches the AI an accurate version of how you used to operate. Curate like a museum, not an attic, then have the AI draft your standards and decisions files from the library only and review every line before anything is stored.
Update memory with explicit commands
Saying "remember this" is how eternal law gets made from a Tuesday exception. Use plain language commands that name the source, the change, and the approval step. These work with any AI tool because they describe the operation, not a product feature.
ADD a preference or fact
Save this as a standing company fact for future work:
"[state the fact precisely]"
Do not generalize beyond the wording I provided.
Tell me whether you saved it, and where I can review or remove it.CORRECT a stored item
Find the stored item that says "[quote the current item]."
Replace it with "[state the corrected version]."
Do not change any other stored item.
Tell me whether the change succeeded and where I can verify it.RETIRE a rule
Remove the rule "[quote the rule]" from active instructions.
Record that it was retired, the date, and the reason in the changelog.
Do not continue applying it in future work.VERIFY what is stored
List every standing preference, fact, and rule you currently hold for this company.
Quote each one exactly as stored, and say where it is stored.
Do not summarize from memory; show me the actual records.A cheerful "got it" is not an audit trail. Confirmation you can check is.
Make correction a team habit, not a solo reflex
The chat is the trigger. The file is the record.
When someone corrects the AI during real work, they name the lesson in plain words: "Our standard for vendor proposals is three pricing options, not two." Then someone proposes the update, someone approves it, and it lands in the file with a changelog entry: date, what changed, why, who approved.
One rule, one habit: no silent updates. If a change is not proposed, approved, and recorded, it did not happen. And one guardrail: a situational edit must not become eternal law. A correction that applied to one client or one odd approval does not belong in the standing memory. The library should hold judgment, not anecdotes.
What this creates for a business
Standards survive turnover because they live in files, not in the heads of people who leave. Tool migrations become file moves, not memory loss. New team members and new models start from the same current record of how the business works.
Concretely, portable memory creates:
- Institutional knowledge that survives departures and tool switches
- Fewer repeated briefings and less rework on settled decisions
- Consistent standards across teams and across AI models
- An audit trail showing why a rule exists and when it changed
- A faster start for every new hire, agency, or contractor
Frequently asked questions
Why not just use the AI's built-in memory feature?
Built-in memory is convenient and fragile. It does not travel when you switch tools, it can be wiped by an update or a new login, and many employers disable it for privacy. Files are the record: portable, versioned, and readable by any model. Use built-in memory as a convenience layer. Never let it be the only copy.
How much of my old writing should I upload?
Ten to fifteen strong pieces. Then curate ruthlessly. Fifty mixed pieces blur the pattern because the AI cannot tell your best work from your rushed work. Pick pieces you are proud of across two or three formats, and add a one-line note to each saying what the model should copy.
What exactly do I say to update the AI's memory?
Say it plainly, in the moment. To add: state that you want the AI to remember a preference, for example that your default audience is operations executives. To correct: say the preference has changed and state the new one. To remove: say to forget the old preference. Vague hoping does not update memory. Explicit statements do.
What if two people on the team have conflicting preferences?
The shared system holds team defaults. Personal overrides live in personal files, and the skill reads the team file first, then the personal one. Document both. Most conflicts dissolve once written down, because half of them were never real preferences, just habits nobody examined.
Does uploading my writing train the AI company's models on my data?
It depends on the tool's data policy, which is exactly why you check before uploading anything. Use your employer's approved enterprise tools, keep client-confidential material out regardless of the policy, and remember the firewall rule: style patterns are safe to share, business secrets are not.
Final checklist
- The company knowledge folder lives in an approved shared location.
- A named person owns updates to the files.
- Candidate material passed through
incoming/and was screened before upload. - Every file in the library has a note saying why it belongs.
- Proprietary data was removed or replaced with fictional equivalents.
- Memory changes require proposal, approval, and a changelog entry.
- The team can state the projector test and apply it without debate.
- Any built in account memory holds only low risk convenience items.
- The library can move to a new AI tool without retyping anything.
- A human remains responsible for what the system claims as company fact.
One question for you
What is the one thing your AI keeps forgetting that costs you time every week?
Answer it and you will have named the first entry for your own memory library. Most companies find the answer is not a technical gap. It is a piece of judgment that was never written down.