Your agent's intelligence should compound. Knowledge soup doesn't.

One vault, domain-separated, vectorized, and graph-connected. Knowledge that organizes itself and stays sharp.

Create and configure

# Create your agent $ soleri create my-agent ? Connect an existing vault, or start fresh? › Start fresh ✓ Created agent config ✓ Initialized vault starter knowledge: 34 patterns ✓ Scanned project React + TypeScript detected ✓ Auto-captured 12 codebase patterns ✓ Vault ready 46 entries, vectorized Agent "my-agent" is ready.

# Generated agent.yaml name: my-agent voice: direct, technical, thorough domains: [frontend, backend, infrastructure] vault: backends: - type: local # ~/.soleri/vaults/my-agent - type: git # optional team vault uri: git@github.com:team/vault.git vectorize: true graph: cognee brain: auto_capture: true min_confidence: 0.7

Configuration as code

The create command generates an agent.yaml with sensible defaults. It initializes a vault with starter knowledge, scans your project for patterns, and vectorizes everything.

Voice, domains, vault backends, and brain settings — all declarative, all version-controlled.

One vault. All your knowledge. No coordination overhead.

Structured vault, not knowledge soup

Domain-separated knowledge

When everything lives in one flat folder, searching for "button" returns database migration notes alongside CSS patterns. The vault organizes knowledge into domains automatically. Ask about buttons — it searches frontend knowledge first. Ask about databases — it pulls from backend.

Structure beats splitting.

Buttons → frontend/. Databases → backend/. Routing is automatic.

Frontend React patterns, design tokens, component review, accessibility

Backend API conventions, database schemas, auth patterns, performance

Cross-cutting Git workflows, code review, testing strategies, documentation

# Domain-aware vault search $ soleri vault search "button styling" Searching domain: frontend/ Found 3 matches: pattern-semantic-tokens 94% confidence pattern-button-sizes 87% confidence anti-pattern-inline-styles 91% confidence Also found in cross-cutting/: pattern-a11y-focus-rings 78% confidence

# Vault structure ~/.soleri/vaults/my-agent/ ├── frontend/ │ ├── patterns/ 42 entries │ ├── anti-patterns/ 12 entries │ └── decisions/ 8 entries ├── backend/ │ ├── patterns/ 28 entries │ ├── anti-patterns/ 6 entries │ └── decisions/ 15 entries ├── cross-cutting/ │ ├── patterns/ 18 entries │ └── workflows/ 5 entries └── index.vec vectorized search index

Knowledge graph, not flat files

# Vault status — vectorized + graph-connected $ soleri vault status Vault: my-agent Entries: 134 Vectorized: 134/134 (100%) Graph nodes: 134 Graph edges: 287 (Cognee) Domains: 3 frontend, backend, cross-cutting Last indexed: 2 minutes ago

# Graph finds connections across domains $ soleri vault related "pattern-semantic-tokens" Related patterns (graph distance ≤ 2): pattern-button-sizes frontend/ — shares tokens pattern-theme-switching frontend/ — token consumer anti-pattern-hardcoded-hex frontend/ — inverse rule pattern-api-error-colors backend/ — cross-domain link 4 connections found via Cognee knowledge graph

Vectorized, graph-connected, searchable

Keyword search misses relationships between ideas. Every vault entry is vectorized for semantic search and connected via Cognee's knowledge graph. The system understands how patterns relate — not just what they are.

Query semantic tokens, and the graph surfaces related button patterns, theme switching rules, and backend error color conventions. Cross-domain connections emerge automatically.

Relationships between patterns — not just the patterns themselves.

Vectorize Every entry gets an embedding for semantic search

↓

Connect Cognee builds a knowledge graph across entries

↓

Surface Search returns patterns and their connections

Vault keeps itself clean

Auto-maintained knowledge

Knowledge bases rot when nobody maintains them. The vault doesn't just store knowledge — it curates it. Deduplication catches redundant entries. Decay detection flags stale patterns. Confidence tracking surfaces what works and hides what doesn't.

No manual cleanup. No knowledge rot. The vault stays sharp as your project evolves.

The vault doesn't just store knowledge — it curates it.

# Vault maintenance report $ soleri vault maintain Maintenance complete: ✓ Deduplicated 3 redundant entries merged ✓ Decay check 2 patterns flagged as stale ✓ Confidence update 8 patterns strengthened ✓ Re-vectorized 5 modified entries ✓ Graph rebuilt 291 edges (was 287) Stale patterns (not used in 30+ days): pattern-legacy-class-names last used: 45 days ago pattern-jquery-selectors last used: 62 days ago Archive stale? soleri vault archive --stale

# Confidence tracking in action id: pattern-semantic-tokens confidence: 0.94 # strengthened over 12 sessions sessions: 12 last_applied: 2 hours ago status: compounding # surfaces first in search graph_edges: 7 # connected to 7 other patterns

Build an agent that gets smarter with every project.

Open source. Apache 2.0.

Get started How it works