Know-Do Graph¶

The Know-Do Graph is MatCreator's long-term knowledge base. It stores what the agent can do, how tasks are performed, what happened in previous sessions, and which observations have become reliable enough to reuse.

Data Model¶

Durable knowledge and working memory are stored together in one SQLite-backed graph. The default location is:

~/.matcreator/.adk/know_do_graph.db

The graph contains entries and typed edges:

Entry type	Meaning
Capability	A reusable skill or ability.
Procedure	A guide, workflow, or multi-step method.
Heuristic	A distilled lesson from repeated successful evidence.
Memory	A session-level observation, result, or intermediate learning.

Runtime Behavior¶

Skills are durable capability entries.
Guides are durable procedure entries.
Agent saves are native memory nodes.
Memory-to-memory relationships use related_memory edges.
Memory linked to a durable skill uses memory_of edges.
Promoted memory links to refined knowledge with refinement_of edges.
Retrieval searches durable entries and unpromoted memory in the same graph.

Progressive Retrieval¶

MatCreator first searches high-level capabilities and procedures. It then conditionally expands into heuristics, constraints, and memories connected to the selected knowledge. This keeps retrieval focused while still allowing the agent to recover detailed lessons when they are relevant.

Review And Distillation¶

The knowledge reviewer can examine graph entries, session memory, and repeated patterns. Successful repeated memories can be promoted into durable heuristics:

matcreator knowledge distill --min-evidence 3

Nodes marked as peer-reviewed or community-tested are protected from mutation during review.

Useful Commands¶

matcreator knowledge seed
matcreator knowledge query "structure generation"
matcreator knowledge search-skills "VASP relaxation"
matcreator knowledge related-skills <start-node>
matcreator knowledge stats
matcreator knowledge distill --min-evidence 3

For legacy database migration, see Knowledge Migration.