Agent Autonomy Primitives

Turn any AI agent into a self-directing worker using five composable primitives: typed memory, task files, project grouping, template schemas, and heartbeat loops.

Prerequisites

CODEBLOCK0

The Five Primitives

1. Typed Memory

Every memory has a type. The type determines where it lives and how it's retrieved.

Type	Directory	When to Use
INLINECODE0	INLINECODE1	Recording a choice with rationale
INLINECODE2

lessons/ | Something learned from experience |
| person | people/ | Contact info, relationship context |
| commitment | commitments/ | Promise made, deliverable owed |
| preference | preferences/ | How someone likes things done |
| fact | inbox/ | Raw information to file later |
| project | projects/ | Workstream with goals and status |

Store with type:
CODEBLOCK1

Rule: If you know WHAT KIND of thing it is, use the right command. Dumping everything into daily notes defeats retrieval later.

2. Task Primitives

A task is a markdown file with YAML frontmatter in tasks/:

CODEBLOCK2

Create tasks:
CODEBLOCK3

Update status:
CODEBLOCK4

Statuses: open → in-progress → done (or blocked)
Priorities: critical > high > medium > INLINECODE22

3. Project Grouping

Projects group related tasks with metadata:

CODEBLOCK5

Tasks reference projects via the project field. Filter tasks by project:
CODEBLOCK6

4. Template Schemas

Templates are YAML schema definitions that control what fields exist on every primitive. They live in templates/ in your vault.

See references/template-customization.md for full customization guide.

Key points:

• - Vault templates override builtins — drop a task.md in templates/ to change the schema
• Add fields (e.g., sprint, effort, client) by editing the template
• Remove fields you don't need
• Change defaults (e.g., default priority = high)
• Validation is advisory — warns but never blocks

5. Heartbeat Loop

The heartbeat is the autonomy mechanism. Wire it into your agent's periodic wake cycle.

Every heartbeat (e.g., every 30 minutes):

CODEBLOCK7

Implementation for OpenClaw agents:

Add to your HEARTBEAT.md:
CODEBLOCK8

For cron-based agents, schedule a recurring job:
CODEBLOCK9

Composing Primitives into Autonomy

The power is in composition, not any single primitive:

CODEBLOCK10

Each cycle compounds:

• - Memory feeds context into task execution (decisions, lessons, preferences inform how work gets done)
• Task execution generates new memories (lessons learned, decisions made, commitments created)
• Lessons improve future execution (mistakes aren't repeated)
• Wiki-links ([[entity-name]]) build a knowledge graph across all files
• Projects provide scope boundaries so the agent doesn't drift

Adapting to Your Setup

See references/adaptation-guide.md for detailed patterns on:

• - Wiring primitives into existing agent frameworks (OpenClaw, LangChain, CrewAI, custom)
• Choosing which primitives to adopt (start minimal, add as needed)
• Multi-agent collaboration through shared vaults
• Migrating from other memory systems

Quick Start: Zero to Autonomous in 5 Minutes

CODEBLOCK11

Anti-Patterns

Don't	Do Instead
Store everything in one big memory file	Use typed memory — decisions/, lessons/, people/
Create tasks without owner/project

Always set --owner and --project | | Ask "what should I work on?" | Read your task queue and decide | | Forget lessons after learning them | clawvault remember lesson immediately | | Skip marking tasks done | Always task done --reason — the ledger tracks transitions | | Create tasks for vague ideas | Put ideas in backlog/, promote to tasks/ when ready | | Modify template schemas constantly | Stabilize schemas early — field renames break existing files |

Obsidian Integration

Because everything is markdown + YAML frontmatter, Obsidian renders your agent's workspace as a human-readable dashboard:

• - Kanban board — open all-tasks.base in Obsidian Bases, drag between status columns
• Blocked view — blocked.base shows what needs human input
• By owner — by-owner.base shows what each agent is working on
• By project — by-project.base scopes views per workstream

The same file is both the agent's data structure AND the human's UI. No sync layer needed.