Vault

Vault organizes agent intelligence into three layers — each serving a different purpose, all working together.

┌─────────────────────────────────────────────────────────────────┐
│  Tier 1: Session State                                         │
│  Redis/MySQL checkpoints — crash recovery, workflow state       │
│  Automatic context window management + auto-summarization       │
├─────────────────────────────────────────────────────────────────┤
│  Tier 2: Working Memory (Surface Memory)                        │
│  Personal scratchpad per agent — facts, insights, preferences   │
│  Deduplication (>80% similarity = update, not duplicate)         │
│  ChromaDB semantic search + keyword fallback                    │
│  Secret masking with {{secret:KEY}} placeholders                │
├─────────────────────────────────────────────────────────────────┤
│  Tier 3: Knowledge Base (Bloqs + Pinecone)                      │
│  Permanent, shared, RAG-indexed documents                       │
│  PDF/CSV/TXT auto-extraction and vectorization                  │
│  Cross-agent knowledge sharing                                  │
└─────────────────────────────────────────────────────────────────┘

Unlike file-based tools (fcontext, Context Layer) or standalone memory servers (Mem0), Vault is fully integrated — memory, secrets, integrations, and RAG search in one platform. No glue code, no separate services to manage.

Most memory tools store flat text strings. Vault stores structured entity relationships — your agents don't just remember facts, they understand your business graph.

Lead: John Doe (#412)
├── Status: Negotiation
├── Notes: "Prefers aggressive outreach on Instagram"
├── Tasks
│   ├── ✓ Scraped 30 competitor profiles
│   ├── ◻ Send proposal PDF (due Friday)
│   └── ◻ Schedule demo call
├── Deliverables
│   ├── website_design.fig (sent)
│   └── brand_guide.pdf (pending)
├── Invoices
│   ├── #1042 — $2,400 (paid)
│   └── #1089 — $4,800 (unpaid)
├── Outreach Steps
│   ├── Step 1: Initial email (completed)
│   ├── Step 2: Follow-up DM (completed)
│   └── Step 3: Phone call (pending)
└── Working Memory
    ├── [fact] Budget is $50k/quarter
    ├── [insight] Opens emails Tuesdays 9am
    └── [preference] Formal tone in proposals

When your sales agent wakes up, it doesn't search for keywords in a text file. One namespace for everything — unstructured memories AND structured CRM entities:

# Unstructured working memory
iris sdk:call memory.search agent_id=11 query="John Doe preferences"
iris sdk:call memory.store agent_id=11 type=fact content="Budget is \$50k"

# Structured entities — same memory.* namespace
iris sdk:call memory.entities agent_id=11              # All leads
iris sdk:call memory.entities agent_id=11 type=tasks   # All tasks
iris sdk:call memory.entities agent_id=11 lead_id=412  # One lead's data
iris sdk:call memory.graph agent_id=11                 # Full relationship map

The result: Your agent knows John's invoice is unpaid, his proposal is due Friday, and he prefers formal tone — without you pasting a single thing into the prompt.

Vault speaks MCP — the protocol Claude, Cursor, and Cline use natively. Install once, context is everywhere.

# Install the Vault MCP context skill
iris marketplace install vault-context

# Or connect directly in your tool's MCP settings
# Server: https://heyiris.io/mcp/vault
# Auth: your IRIS API key

Once connected:

• Claude Code pulls agent memory and document context automatically
• Cursor accesses your project docs, contracts, and research without paste
• Cline and any MCP-compatible agent gets your full knowledge base

No glue code. No re-uploading. Upload once — every tool uses it.