Memory and skills

Workspace memory, Dreams, and how skills load and self-learn. For the concepts, read Memory & Dreams and Skills & Self-Learning.

Memory

Field	Description	Default
Memory.Enabled	Lets new sessions use and maintain workspace memory. Dreams and welcome suggestions also require it	true
ConsolidationModel	Model for Dreams runs. Empty uses the main model	Empty

Skills

Field	Description	Default
Skills.DisabledSkills	Skill names disabled for this workspace. A disabled skill stays on disk but is left out of agent context	[]
Skills.IncludeSharedSkills	Also loads personal skills from ~/.agents/skills, the directory shared with other agent tools. Takes effect on restart	true
Skills.IncludeBuiltInSkills	Deploys the skills DotCraft ships, such as dotcraft-guide, into the workspace on start. A host that brings its own skills turns this off. Takes effect on restart	true
Skills.SelfLearning.Enabled	Master switch for agent skill self-learning; off hides skill editing from the model	true
Skills.SelfLearning.VariantMode	Skill variant write mode: enabled routes self-learning updates to workspace-local skill variants, disabled turns variants off	enabled
Skills.SelfLearning.MaxSkillContentChars	Max chars for a single SKILL.md written through self-learning	100000
Skills.SelfLearning.MaxSupportingFileBytes	Max bytes for a single supporting file written through self-learning	1048576

Self-learning example:

{
  "Skills": {
    "SelfLearning": {
      "Enabled": true,
      "MaxSkillContentChars": 100000,
      "MaxSupportingFileBytes": 1048576
    }
  }
}

SkillManage(action, ...) reference:

Action	Required parameters	Purpose
create	name, content	Create a new workspace skill
patch	name, oldString, newString	Local patch of SKILL.md or supporting file
edit	name, content	Replace an existing workspace skill's SKILL.md
write_file	name, filePath, fileContent	Write a supporting file
remove_file	name, filePath	Delete a supporting file

create triggers a kind: skill approval, and destructive deletes require approval too. Self-learning writes only to the current workspace's skill directory. System and personal skills are read-only, supporting files may only live under scripts/ or assets/, and absolute paths or .. traversal are rejected.