Beliefs, Personas, and Memory

Mind cards can contribute three categories of cognitive content that are injected into the agent's system prompt at write time. Each type is declared in the card manifest and scaffolded via drwn card source subcommands.

Beliefs

A belief is a file of factual assertions — domain truths the author wants the agent to treat as given. Beliefs are composed into the system prompt as read-only context.

On disk (inside a card source): beliefs/<entry>/BELIEF.md

# Scaffold a new belief entry
drwn card source add-belief @your-handle/mind engineering --visibility public

# Remove a belief entry (and its files by default)
drwn card source remove-belief @your-handle/mind engineering
drwn card source remove-belief @your-handle/mind engineering --keep-files

The --visibility flag is required and controls which consumers receive the belief:

Value	Meaning
`private`	Only the card author's projects
`internal`	Org-internal consumers
`public`	Any consumer

In the card manifest (card.json), beliefs appear under beliefs.include:

{
  "beliefs": {
    "include": ["engineering"]
  }
}

Personas

A persona is a file that shapes the agent's behavioral and stylistic identity — tone, voice, scope of engagement, communication style. Only one persona is composed into the system prompt per write; when multiple mind cards contribute personas, the active stack order determines precedence.

On disk: persona/<entry>/PERSONA.md

# Scaffold a persona entry
drwn card source add-persona @your-handle/mind voice --visibility internal

# Remove a persona entry
drwn card source remove-persona @your-handle/mind voice
drwn card source remove-persona @your-handle/mind voice --keep-files

In the card manifest:

{
  "persona": {
    "include": ["voice"]
  }
}

Memory

Memory is layered structured knowledge that accumulates over time. Three layers are supported, each designed for a different retention horizon:

Layer	Path	Typical content	Format
L4	`memory/l4/<entry>/`	Short-term project context, recent decisions	`md`
L5	`memory/l5/<entry>/`	Mid-term accumulated patterns, team knowledge	`jsonl` or `mixed`
L6	`memory/l6/<entry>/`	Long-term durable facts, stable domain knowledge	`jsonl`

# Scaffold an L4 markdown memory entry
drwn card source add-memory @your-handle/mind context --layer l4 --visibility private --format md

# Scaffold an L6 JSONL memory entry
drwn card source add-memory @your-handle/mind raw --layer l6 --visibility private --format jsonl

# Remove a memory entry
drwn card source remove-memory @your-handle/mind context --layer l4
drwn card source remove-memory @your-handle/mind context --layer l4 --keep-files

Flags for add-memory:

Flag	Required	Values	Default
`--layer`	yes	`l4`, `l5`, `l6`	—
`--visibility`	yes	`private`, `internal`, `public`	—
`--format`	no	`md`, `jsonl`, `mixed`	`md`

In the card manifest, memory entries appear under the layer they belong to:

{
  "memory": {
    "l4": { "include": ["context"] },
    "l6": { "include": ["raw"], "format": "jsonl" }
  }
}

Visibility and composition

All three content types carry explicit visibility metadata. drwn write filters content based on visibility and the current project's trust policy. Content marked private is never exported when the card is published to a catalog.

When multiple mind cards are active, content from all active cards is composed in active-stack order. See Minds for the composition model.

Beliefs​

Personas​

Memory​

Visibility and composition​

See also​

Beliefs

Personas

Memory

Visibility and composition

See also