ClawVault

Structured memory for AI agents. Typed markdown primitives that compound over time.

Every memory is a markdown file with YAML frontmatter — a task, a decision, a person, a lesson — each following a schema defined in templates/. The agent reads and writes these files. The human browses them in Obsidian. No database. No vendor lock-in. Just files.

Requirements

• Node.js 18+
• qmd installed and on PATH (hybrid BM25 + vector search)

Install

As an OpenClaw Plugin (recommended)

openclaw plugins install clawvault

This installs ClawVault as a memory plugin. It replaces OpenClaw's built-in memory with:

• Auto-recall — injects relevant memories before each agent turn
• Auto-capture — observes conversations and stores durable knowledge automatically
• Session recap — on wake, provides context from active tasks, recent decisions, and preferences
• 4 tools — memory_search, memory_store, memory_get, memory_forget

After install, configure the vault path:

openclaw config set plugins.clawvault.config.vaultPath ~/my-vault

As a Standalone CLI

npm install -g clawvault

Quick Start

# Initialize a new vault
clawvault init ~/my-vault --name my-brain

# Set up Obsidian Bases views (tasks, projects, backlog)
clawvault setup

# Check vault health
clawvault doctor

# Search your vault
clawvault search "deployment decision"

How It Works

Typed Primitives

Every piece of memory has a type defined by a template:

---
primitive: task
fields:
  status:
    type: string
    required: true
    default: open
    enum: [open, in-progress, blocked, done]
  priority:
    type: string
    enum: [critical, high, medium, low]
  owner:
    type: string
  due:
    type: date
---

Default templates: task, decision, lesson, person, project, checkpoint, handoff, daily, trigger, run, party, workspace.

Malleable Schemas

Don't like the defaults? Drop your own template in your vault's templates/ directory. Add fields, remove fields, create entirely new types. The plugin reads YOUR schemas, not ours.

Hybrid Search

ClawVault uses qmd for search — BM25 keyword matching combined with vector similarity and reranking. Entirely local. No API keys needed.

Obsidian Integration

Your vault IS an Obsidian vault. Tasks become Kanban boards. Decisions are searchable. Wiki-links build a knowledge graph. Five generated Bases views out of the box:

• All tasks
• Blocked items
• By project
• By owner
• Backlog

CLI Commands

Core

Command	Description
init [path]	Initialize a new vault
setup	Auto-discover and configure a vault, create Obsidian views
store	Store a new typed memory document
capture <note>	Quick-capture a note to inbox
doctor	Diagnose vault health

Search & Context

Command	Description
search <query>	BM25 keyword search via qmd
vsearch <query>	Semantic vector search via qmd
context <task>	Generate task-relevant context
inject <message>	Inject relevant rules and decisions

Session Lifecycle

Command	Description
wake	Start a session (recover + recap)
sleep <summary>	End a session with a handoff
checkpoint	Save state for context-death resilience
recover	Check for and recover from context death

Observation Pipeline

Command	Description
observe	Process sessions into observational memory
reflect	Promote observations to weekly reflections
reweave	Backward consolidation — mark superseded observations

Tasks & Projects

Command	Description
task	Task management (create, list, update, transition)
project	Project management
kanban	Kanban board view
status	Vault health and statistics

Utilities

Command	Description
template	Manage document templates
graph	Show typed memory graph summary
entities	List all linkable entities
link [file]	Auto-link entity mentions
compat	Check OpenClaw compatibility
embed	Run qmd embedding for pending documents

Architecture

     HUMAN (Obsidian)
     Browse, edit, approve
           │
           ▼
  ┌─── VAULT (markdown) ───┐
  │  Typed primitives       │
  │  Knowledge graph        │
  │  Template schemas       │
  └───┬──────────────┬──────┘
      │              │
 AGENT (Plugin)   CLI (Developer)
 Auto-capture     Direct CRUD
 Auto-recall      Search, graph
 Session recap    Tasks, projects

OpenClaw Plugin Details

The plugin hooks into the OpenClaw lifecycle:

• before_agent_start — auto-recall: searches vault for context relevant to the current conversation and injects it
• message_received — auto-capture: observes incoming messages for durable information worth storing
• agent_end — captures any final observations from the agent's response
• before_compaction — preserves important context before conversation compaction

Configuration in openclaw.plugin.json:

Option	Default	Description
vaultPath	—	Path to vault directory
collection	clawvault	qmd search collection name
autoRecall	true	Inject memories before each turn
autoCapture	true	Auto-store from conversations
recallLimit	5	Max memories per recall

What Compounds

• Decisions accumulate into institutional knowledge
• Lessons prevent repeated mistakes
• Tasks with transition ledgers track how work happened
• Projects group related work across hundreds of sessions
• Wiki-links build a knowledge graph that grows richer over time

The agent that runs for a year generates compounding value. Every lesson stored makes the next task cheaper.

License

MIT