DM is shared context for human + AI teams. It prepares what applies to each task — your team’s Rules, Skills, Decisions, and repository context — and brings what people and agents learn back for review.
A human and a connected agent both use a DM workspace. The human approves Rules, Skills, and Decisions before they become trusted. The agent can propose new learning for review, and can read already-approved context that applies to the work. Repository intelligence is available in the workspace. Memory is in progress and is not automatically trusted. Connecting an agent is explicit.
Memory / context · Rule · Skill · Decision
Context breaks in both directions: what your team knows doesn't reliably reach the work, and what the work teaches you doesn't reliably come back.
Decisions, rationale, constraints, and exceptions live in meetings, chats, documents, tickets, and conversation — not in the next prompt.
Decisions get replaced. A rule may apply only in one place. Assumptions expire. Yesterday’s memory can be wrong today.
Useful context stays trapped in a session, a chat, or one person’s head instead of becoming something the team can reuse.
Work produces discoveries that rarely come back as shared context. The next person — or the next agent — reconstructs them.
Connected agents read already-approved context that applies — no new approval on every read. New learning comes back as a proposal; people decide what becomes trusted.
Real DM UI · synthetic demo workspace. No customer data.
Remembering what happened is useful. Persistent work also needs the system around the task, and what the team has already decided.
Ready in a DM workspace today: Rules, Skills, Decisions, repository intelligence, architecture context, Prepare Context, and MCP access. Memory is in progress.
These complement each other in one product: what your team decided, how the system fits together, and — when it is available — what happened before.
More context is not always better.
The work starts without what the team already knows.
A dump of memory and files hides what actually matters.
Replaced decisions and expired assumptions still look current.
Context the person or agent should not use still shows up.
Current, relevant, shared, approved, and permitted for this work.
A developer no longer works with one assistant. Claude Code for implementation. Cursor or Copilot in the editor. Another AI tool for a specific job. Each capable. None reliably carries forward what the others established.
Same task. Implement tenant file upload.
Nothing here was a coding mistake. The code compiled, the tests passed, and every line broke something the team had already settled. The organisation already knew. DM made that knowledge available to the work.
The agent has the ticket and the repo. It misses an architectural decision, repeats a rejected approach, and loses useful context from the last session.
Prepare Context can include applicable Rules, relevant Decisions, useful Skills, and repository intelligence. Memory is added when it is available — not as a finished composed pipeline.
We already run this on our own engineering work, over MCP, with Cursor, Claude Code, Copilot CLI, and other MCP clients. Connecting an agent is explicit.
Using coding agents? Try DM on one recurring task Coding-agent alignment
The records behind the loop. Each one carries where it came from, where it applies, who approved it, and whether it is still current — which is what makes it safe to hand to a connected agent.
What must, must not, or needs approval. The boundaries the team already agreed to — carried to the work instead of hoped for.
The method for a class of work, so it is not reinvented per task — and not re-explained to every new agent and every new teammate.
What was chosen, why, where it applies, and what it replaced. The durable object behind the rule — and the one thing documentation reliably loses.
Memory vs a decision
Memory records what happened. A Decision is what is still current — including what replaced an older choice.
Memory “We used configuration X.”
“Configuration X is deprecated. Configuration Y is current for Project A.”
Memory records what happened. A Decision is what is still current.
Memory is in progress in the same workspace — including memory your agents already use, when connected.
Coding agents make the gap obvious. DM is for any human + AI team that needs shared context to last longer than a session.
Keep constraints and methods with the people and agents shipping the work.
Carry forward why the system is this way — not only what the repo looks like today.
What was chosen, what was rejected, and whether it is still current.
The next person should not reconstruct months of “why” from tickets and chat.
Several capable tools on one persistent project — without starting from scratch each session.
Humans and connected agents sharing one workspace of what is current.
Give your human + AI team one place for the context behind the work — and let the next person or connected agent pick up with less reconstruction.
Only the context a person or connected agent is allowed to use. Trust & security