Memory

The plugin keeps a Memory folder in your vault for notes that shape how Steward remember its previous works.

Tool instructions memory

Extra guidelines per tool that are merged into the agent prompt when those tools are active.

The Memory notes are meant for the model (and you when you want to tune behavior). Steward reads validated YAML blocks from the file; prose above or between blocks is for the model only and is not parsed as guidelines.

See YAML blocks in markdown for the shared pattern (fences, validation, and the status field).

Location

Item	Path
Folder	Steward/Memory/
Agent guide	Steward/Memory/Agent.md — how to edit tool instructions (for the model)
Tool instructions	Steward/Memory/Tool instructions.md

On first run, Steward creates the folder and starter Agent.md and Tool instructions.md if they do not exist. Saving or editing Tool instructions triggers validation; frontmatter status is updated automatically (Valid or Invalid: …).

Agent.md

Agent.md is a short, model-facing checklist: section layout (heading per tool, no underscores in the heading; YAML block below), the tool_instruction field table, and re-read Tool instructions after edits to confirm status is valid.

How tool instructions work

1. Built-in guidelines: Each tool already has descriptions and guidelines in Steward. Those are always the baseline.
2. Memory guidelines: Valid tool_instruction YAML blocks in Tool instructions.md add more bullet points for the matching tool.
3. When they apply: Memory guidelines are merged into the system prompt for active tools only (same timing as Guardrails rule text for tools). They apply on the next agent turn after the file is loaded or updated.

Hard-coded guidelines appear directly under each tool heading in Tool → Guidelines; Guardrails and Memory use their own subheadings when present.

Block format

Each guideline set is a ```yaml fence with name: tool_instruction:

Field	Required	Description
name	Yes	Must be tool_instruction
tool	Yes	Tool id (e.g. shell, grep, edit), must match a Steward tool name
enabled	No	false skips this block. Default true
guidelines	Yes	List of strings; one guideline per line

The whole note can be disabled with enabled: false in frontmatter (same pattern as guardrails and skills).

Example block

name: tool_instruction
tool: shell
enabled: true
guidelines:
  - (Steward adds its custom guidelines here)

You or Steward can add multiple blocks for different tools, or several blocks for the same tool (guidelines are appended in file order).

Only tool_instruction blocks are loaded. Other YAML fence types in the file are reported as validation errors.

Example: RTK and the shell tool

RTK is a CLI proxy that compresses common command output (git, tests, docker, etc.) to save tokens in LLM context. Steward does not hard-code RTK. You (or the model) record how to use RTK in Tool instructions for the shell tool.

Scenario

You ask Steward to help save tokens when running shell commands, and to install RTK if needed:

Install rtk for token-saving on shell output, and remember to use it for routine commands.

Steward may install RTK (e.g. via shell after your confirmation) and update Steward/Memory/Tool instructions.md with a tool_instruction block for shell, similar to:

name: tool_instruction
tool: shell
enabled: true
guidelines:
  - Run routine non-interactive commands as rtk <command> (e.g. rtk git status) so output stays compact for follow-up turns.
  - Use plain shell (no rtk prefix) for interactive programs. output.

What happens on later turns

1. The model activates shell (or it is already active) via activate_tools.
2. The activate result reminds the model that supplemental guidelines can be edited in Steward/Memory/Tool instructions.md.
3. When the model calls the shell tool, Steward shows your command for confirmation as usual; Memory does not rewrite commands automatically.
4. On each agent turn, the shell entry under Tool → Guidelines lists built-in bullets first, then Guardrails and Memory subsections when present; Memory lines come from Tool instructions.md, so the model tends to choose rtk git status instead of raw git status without you repeating the preference every chat.

Transcript and interactive shell behavior are described in Shell command & Terminal. Memory only affects instructions in context, not automatic command wrapping.