Where should context live?

I have been using Claude's memory for a while now, and it is the best working relationship I have had with a tool. It remembers how I like things written, which projects are half done, what I decided last week and why. I do not have to re-explain myself. The agent just picks up where we left off.

The trouble starts the moment someone else needs the same thing.

The problem with personal memory

Claude's memory is built around one person. It learns your habits, your preferences, your shortcuts. That is exactly why it works so well. It is also why it does not travel. A teammate cannot inherit what my agent has learned, and the decisions that matter to the whole team end up sitting in one person's conversation history.

That is a quiet kind of risk. If I leave, or simply go on holiday, the context goes with me. If a customer asks why we chose a certain price, the answer is somewhere in a chat I had three weeks ago, and nobody else can find it.

The problem with a shared database

The obvious fix is to put everything in the company system. Write it down, make it searchable, let everyone read it. But a shared database has its own failure mode. It goes stale. People trust what is written, so an outdated page does more damage than no page at all. And it is slow to keep up with the way people actually think and work, which is messy, fast, and full of half-formed ideas.

Split it by what the context is

I have come to think the question is not which one wins. It is what kind of context you are holding.

• Personal context belongs to the individual: how someone likes to work, their drafting style, the threads they are juggling. This is where Claude's memory earns its keep. Keep it there.
• Shared context belongs to the team: decisions, customer facts, product reasons, the "why we did it this way." This needs an owner, a place anyone can read, and a way to be corrected.

The mistake is trying to force one layer to do the other's job. A personal memory cannot be a company wiki. A company wiki cannot learn how you think.

Two rules that make the split work

First, decide which one wins when they disagree. If the agent remembers that we dropped a feature, but the shared record says we still ship it, the shared record should win, and the agent should be pointed at it. Otherwise the most confident source takes over, and that is usually the wrong one.

Second, make the shared layer easy to write to. If adding a decision takes ten minutes of formatting, it will not happen. If it takes one sentence from the agent in the middle of a conversation, it will.

Where I have landed

Let the agent hold the personal layer and let the team hold the shared one. Give the agent a clear path into the shared layer so that what it learns for one person can become something the whole team can use. The context should live where the people who need it can actually reach it.

That is the question worth answering before you start building. Not "which tool," but "who needs to know this, and how long should they be able to trust it?"