AI can be remarkably capable inside one conversation and surprisingly forgetful at the start of the next one.
I keep finding the useful work trapped in the chat: a decision, a corrected assumption, a procedure, a list of open questions, or a better way to run the next analysis. The conversation ends, and the durable value becomes something I have to remember to reconstruct.
I am designing a vault to give that work a home. The first version is deliberately simple: Markdown files, folders, links, and a small set of instructions that tell an assistant how to load context and leave useful results behind.
A memory that belongs to the work
The core idea is that context should belong to the project rather than the conversation. A project needs a durable brief that explains what it is, why it matters, what is currently true, and where the important supporting material lives.
That changes the start of a session. Instead of asking me to repeat the history, an assistant can read the current project record. Instead of treating the latest chat as truth, it can see decisions and status in the same place the next assistant will see them.
Markdown is a good starting medium because it is plain, portable, searchable, and readable without a special application. The structure can evolve without locking the knowledge inside one model provider or one software product.
A simple operating shape
Each active project gets a small set of dependable records. The project brief is the hub. Tasks show the open work. Decisions explain the calls that should not be relitigated. A status log shows movement over time. Meetings and reference material link back to the project they affect.
Separate from the project content, a shared skills area holds reusable ways of doing work. If a reporting process, research method, or document pattern will matter again, it should become a named procedure rather than a memory inside one transcript.
The folder structure is not the product. Its job is to make the next action unsurprising. When an assistant finishes a piece of work, there should be an obvious answer to the question: where does the durable result belong?
Instructions close to the work
A general assistant prompt cannot carry the details of every project. The vault lets instructions sit closer to the thing they govern. Global rules can define how the workspace operates, while a project brief can carry the local constraints, language, and sources that make its work distinct.
The assistant should load the relevant material before acting. That means reading the shared workspace rules, locating the project, checking its current status, and following links only as far as the task requires.
This is partly about accuracy, but it is also about consistency. If different assistants can enter through the same project record, the work becomes less dependent on which conversation happens to be open.
Write-back is the real test
Reading the vault is only half the design. If an assistant uses the context but leaves the result in chat, the next session is not actually smarter.
The completion test should be whether the project changed. A finished task is checked. A new decision is recorded with its reasoning. A status change reaches the log. A reusable method becomes a procedure. A correction updates the source that would otherwise keep producing the old answer.
This write-back loop is what turns a folder of notes into a working memory. The value is not that the system can store more. It is that each useful session can improve the starting point for the next one.
What I am trying to learn
I do not know yet how much structure is enough. Too little and the vault becomes a pile of notes. Too much and maintaining the system becomes its own job.
The working bet is that a few stable project records, strong links, reusable skills, and a clear read-work-write-back rhythm can carry most of the value. The structure should grow only when real work exposes a repeated need.
What I want to see is simple: can a new session understand the project faster, repeat fewer mistakes, and leave behind something the next session can use? If that happens consistently, the vault will be more than storage. It will be the beginning of continuity for AI-assisted work.