Plan Memory

PRD authoring and plan execution lifecycle with tool restrictions during planning.

Overview

Plan Memory manages a structured planning workflow: the agent enters a restricted "plan mode" where only read-only tools are allowed, writes a PRD document, structures a PlanNode execution tree, then exits to execute the plan with full tool access.

Slot: 240 (Slot.PROCEDURAL - 10)
Scope: thread (default)
Default budget: { min: 100, max: 3000 }

Usage

import { planMemory } from '@noetic/core';

const layer = planMemory();

The layer is included by default in the Noetic CLI. Users type /plan to enter plan mode.

Configuration

interface PlanMemoryConfig {
  scope?: MemoryScope;
  additionalAllowedTools?: string[];
  maxPrdLength?: number;
  maxTreeDepth?: number;
}

Field	Type	Default	Purpose
`scope`	`MemoryScope`	`'thread'`	Persistence scope
`additionalAllowedTools`	`string[]`	--	Extra tools allowed during plan mode
`maxPrdLength`	`number`	`50000`	Maximum PRD content length
`maxTreeDepth`	`number`	`5`	Maximum plan tree nesting depth

Phase Lifecycle

The layer manages a state machine with five phases:

idle — No active plan. recall returns nothing.
planning — Read-only tools only. LLM writes PRD and plan tree.
executing — Full tool access. Plan context injected into each LLM call.
completed — Execution succeeded. Outcome recorded.
failed — Execution failed or was aborted. Outcome recorded.

LLM Tools

The layer exposes four LLM tools and one data projection via its provides map:

Tool	Description
`plan/enterPlanMode`	Enter plan mode. Accepts optional `goal` string.
`plan/updatePrd`	Replace PRD content with new markdown.
`plan/setPlanTree`	Set the `PlanNode` execution tree.
`plan/exitPlanMode`	Exit plan mode (`execute` or `cancel`).
`status` (data)	Read-only: `{ phase, hasPrd, hasPlanTree, version }`

Tool Restrictions

During the planning phase, beforeToolCall restricts tool usage:

Allowed: Read, Grep, Find, Ls, activateSkill, and all plan/* tools
Denied: All other tools (Write, Edit, Bash, etc.)

Additional tools can be allowed via additionalAllowedTools in the config.

Integration with compilePlan

The plan layer stores the PlanNode tree produced by the LLM. For programmatic execution, extract the tree and pass it to compilePlan():

import { compilePlan } from '@noetic/core';

const compiled = compilePlan(planTree, agents, constraints);

The default CLI flow uses context injection instead — the plan is recalled into the LLM's view and the model executes by making tool calls directly.