3. Memory for AI (M4AI)
Memory for AI (M4AI) lets you build AI agents that work with the knowledge graph in DCH. DCH remains the source of truth: agents read the parts of the graph you give them access to, and every agent, memory and setting is versioned, so you always know what an agent was working with.
You find M4AI in the sidebar under M4AI:
- Agents: the AI agents themselves, with their instructions, knowledge, skills, tools and model.
- Analysis Engines: sandboxed engines that analyse files you hand them.
- Library: the building blocks agents use.
- Model & runtime configs: which model an agent runs on, and its budgets.
- Memories: the parts of the knowledge graph an agent may read.
- Skills: reusable instructions an agent loads when it needs them.
- Sessions: conversations with your agents and analysis engines.
How the parts fit together
A typical setup takes these steps:
- Create a Model & runtime config for the model you want to use.
- Create a Memory over the target entities or views and functions the agent needs, and publish it once it is ready.
- Optionally, write and publish Skills.
- Create an Agent, give it instructions, attach the memory, skills and tools, pick the model config, and publish it.
- Run the agent and talk to it in Sessions.
Versions
Agents, memories and skills have draft and published versions. You edit a draft; publishing locks it. A published agent version pins the exact versions of everything it uses: memories, skills, child agents and the model config revision. Changing one of them later does not change the agent until you create a new draft and publish again.
Model & runtime configs and Analysis Engine profiles work with revisions instead: every save creates a new revision, and agents and profiles keep the revision they pinned.
Access
M4AI requires a license that includes it. Without one, the M4AI entries in the sidebar are locked.
The first generation of M4AI is described under Legacy: M4AI v1.