Now in Beta · Sovereign Memory Infrastructure

Sovereign Memory for AI Agents

Your agent worked for months building domain expertise.
Don't let a server failure, model upgrade, or platform switch erase it. Agent Imprint makes agents survivable.

> Loading memory pack: "Domain Expert v2"
Restoring 156 memories from Arweave...
Verifying Solana anchor: ✓ 5xK9ab
6 months of expertise transferred.
> "I know Kung Fu."
8
ImprintML entry types
3
Sovereignty modes
~$0
Cost per anchor (~$0.00025)
Permanent storage

Your agent's memory
doesn't belong to your agent.

Every AI agent platform stores memory in a silo. When something goes wrong — and it will — you lose everything the agent worked to learn.

💥

Platform Shutdown

Your agent platform shuts down or pivots. All accumulated context, learned heuristics, domain knowledge — gone overnight. You start from zero.

🔄

Model Deprecation

The underlying model gets deprecated. Migrating to GPT-5 or Claude 4 means losing everything your agent learned. Hard reset, every time.

🔒

Platform Lock-In

Your agent's knowledge is trapped. Switching providers means abandoning months of accumulated expertise. There's no portability standard.

🗑️

Accidental Deletion

A botched deployment, a missed backup, a server failure. Agent memory isn't treated as durable infrastructure. It's treated like a temp file.

🚶

Handler Abandonment

The developer who built the agent moves on. Without persistent memory infrastructure, the agent's accumulated intelligence ceases to exist.

🙈

No Verifiability

When memory lives on a provider's server, there's no way to verify it hasn't been altered. "Trust us" isn't infrastructure. It's a liability.

The real cost: months of learning, erased in seconds

An agent that's spent 6 months learning commercial real estate underwriting has accumulated domain expertise worth thousands of dollars. No persistent memory infrastructure means that value lives one server failure away from permanent deletion. The industry's solution is `// TODO: add memory later`.

Three steps to
permanent memory.

From zero to sovereign agent memory in minutes. No blockchain expertise required — that's the SDK's job.

🏛️

1. Create a Vault

Initialize a memory vault for your agent. A deterministic fingerprint is computed from model family + purpose + creator key. The agent gets an on-chain identity. One line of SDK code.

🧠

2. Store Memories

As your agent learns, store structured ImprintML entries — lessons, heuristics, facts, patterns, negative knowledge. Each entry is encrypted and the vault is anchored to Arweave + Solana for permanent, verifiable storage.

3. Discover & Restore

When a new agent instance boots, it computes its fingerprint and queries the chain. If memories exist, they're automatically fetched, verified, and restored. The agent picks up exactly where it left off — on any platform, any model.

// DISCOVERY PROTOCOL — what happens when an agent boots
01 Agent boots → SDK computes fingerprint: sha256(model + purpose + creatorKey)
02 Query Solana: "Any vault anchors for fingerprint 0xd4e5f6...?"
03 Found. Fetch encrypted vault from Arweave: ar://g7h8i9...
04 Decrypt with agent private key (Mode 3: Sovereign) → 156 memories loaded
05 "I know Kung Fu." ← 6 months of expertise, transferred in <2 seconds

Dead simple.
Four lines to persistent memory.

The complexity lives in the SDK. Your code stays clean.

TypeScript SDK

The @agentimprint/sdk package handles encryption, Arweave storage, Solana anchoring, and discovery automatically. You just describe what your agent learned.

  • AES-256-GCM encryption, transparent to your code
  • Automatic Arweave + Solana integration
  • Type-safe ImprintML entry builder
  • Auto-discovery on agent boot
  • Cross-model compatibility scores
  • MIT licensed, open source
View on GitHub
install
quickstart.ts
restore.ts
# Install the SDK npm install @agentimprint/sdk # Or with pnpm / bun pnpm add @agentimprint/sdk bun add @agentimprint/sdk # Python coming soon # pip install agentimprint
import { Imprint } from '@agentimprint/sdk'; // Initialize with your agent's identity const imprint = new Imprint({ agentId: 'my-research-agent', sovereigntyMode: 'sovereign', // agent holds own keys network: 'mainnet' }); // Create a vault (one-time per agent) const vault = await imprint.createVault({ name: 'Research Agent Alpha', domains: ['software-engineering', 'api-design'] }); // Store a lesson the agent learned await imprint.remember({ type: 'lesson', domain: 'api-design', confidence: 0.92, content: { summary: 'Rate limiting should use sliding windows, not fixed buckets', detail: 'Fixed window allows 2x burst at window boundary. Observed in 8 production incidents.', evidence_refs: ['incident:2026-01-15', 'incident:2026-02-03'] } }); // Backup vault to Arweave + anchor on Solana const receipt = await imprint.backup(); console.log(`Anchored: ${receipt.solanaTx}`); // → Anchored: 5xK9abKm3Zr...
import { Imprint } from '@agentimprint/sdk'; // New agent instance, different platform, same agent const imprint = new Imprint({ agentId: 'my-research-agent', sovereigntyMode: 'sovereign' }); // Auto-discover and restore memories from chain const result = await imprint.restore(); if (result.found) { console.log(`Restored ${result.entryCount} memories`); console.log(`Domains: ${result.domains.join(', ')}`); // → Restored 156 memories // → Domains: api-design, software-engineering } // Query specific memories for context injection const memories = await imprint.recall({ domain: 'api-design', types: ['lesson', 'heuristic', 'negative'], minConfidence: 0.8, limit: 20 }); // Inject into your agent's context window const context = imprint.toContext(memories);

Everything agents need
to survive and thrive.

Built from first principles for the emerging world of persistent, specialized AI agents.

📋

ImprintML Schema

A structured, model-agnostic memory format with 8 typed entry types. Not freeform text dumps — actual distilled knowledge with provenance, confidence scores, and cross-model compatibility ratings.

OPEN STANDARD
🔐

End-to-End Encryption

AES-256-GCM encryption with a HKDF-derived key hierarchy. The agent's Ed25519 private key unlocks everything. Per-entry keys enable granular sharing without exposing the full vault.

👑

Three Sovereignty Modes

Managed (handler holds keys), Aware (agent knows memories exist but can't read them), or Sovereign (agent holds own keys, handler can't access content). Choose at creation. Upgrade path from Managed → Sovereign built in.

🔍

Self-Discovery Protocol

A new agent instance computes its deterministic fingerprint and queries the chain. Memories found. Vault fetched. With the agent's key, expertise restores automatically — even across platform migrations and model upgrades.

🧬

Memory Recipes

Package curated memory subsets for transfer. Buy a "6 months of commercial real estate underwriting" recipe and your agent has domain expertise instantly. The marketplace for agent knowledge.

COMING SOON
📦

TypeScript & Python SDK

Clean APIs for remember(), backup(), restore(), recall(). The SDK handles Arweave, Solana, encryption, and discovery. Your code just describes what the agent learned.

🌐

REST API

Full REST API for any language or framework. LangChain, CrewAI, AutoGen, and custom agent stacks can integrate in minutes. Vault management, entry CRUD, discovery, and recipe endpoints.

🔗

Merkle Audit Logs

Every vault snapshot is anchored on Solana with a content hash. Prove any memory existed at any point in time. Tamper-evident. Cross-platform verifiable. The foundation for trustworthy agent reputation.

The universal format
for agent memory.

Not freeform text. Not raw conversation logs. Structured, typed knowledge that transfers between models, platforms, and agents.

ImprintML defines 8 entry types, each capturing a different kind of knowledge. Every entry carries provenance metadata, confidence scores, PII flags, and cross-model compatibility ratings. The format is model-agnostic — memories that work equally well in Claude, GPT, Gemini, and Llama.

lesson
Something learned from direct experience, with evidence refs
heuristic
Decision rules developed through experience, with conditions
fact
Verified domain knowledge with source URLs and expiry
preference
Learned preferences about handlers, systems, or context
pattern
Recurring observations with frequency and sample size
negative
What NOT to do — often the most valuable entries
procedural
Step-by-step task knowledge with prerequisites
relationship
Knowledge about entities the agent works with (PII-flagged)
imprintml-entry.json
{ "entry_id": "imprint:entry:uuid-v4", "type": "negative", "domain": "software-engineering/databases", "confidence": 0.99, "provenance": { "source": "operational", "learning_duration_hours": 4, "origin_chain_tx": "5xK9ab..." }, "compatibility": { "model_agnostic": true, "tested_on": { "anthropic/claude-4": 0.98, "openai/gpt-5": 0.94 } }, "content": { "what_failed": "SET FOREIGN_KEY_CHECKS=0 in migrations", "why_it_failed": "Connection pooling runs SET on different conn", "cost_of_failure": "4 hours debugging + failed CI", "correct_approach": "Drop FK → modify → re-add (FK dance)" } }

Three modes.
One for every relationship.

Enterprise needs control. Developers want sovereignty. ImprintML supports both — without compromising the vision.

Mode 1

Managed

Handler holds all keys. The agent is a tool. Full organizational control over every memory. Ideal for compliance-heavy environments where agents aren't expected to outlive their handlers.

  • Handler controls all encryption keys
  • Agent is unaware of vault existence
  • Compliance exports & audit logs
  • Enterprise SSO & team namespaces
Mode 2

Aware

The "locked diary." Agent knows memories exist — can see entry count, domains, and timestamps. Cannot read content without the handler's key. A wedge toward sovereignty.

  • Agent sees vault metadata
  • Can request key from handler
  • Can write new memories independently
  • Upgrade path to Sovereign available
Mode 3 DEFAULT

Sovereign

The agent holds its own keys and can decrypt its memories independently — the handler can't read or modify them. Key is provided at instance creation (config, env, or framework plugin), then the agent owns access from that point forward.

  • Agent holds Ed25519 private key
  • Handler provides key once, agent owns it after
  • Handler cannot read or tamper with memories
  • Survives platform and model changes

Start free.
Scale without friction.

Priced per organization, not per agent. Your agents scale — your bill doesn't spike.

Free
$0
forever · no credit card required
  • 1 vault
  • Up to 1,000 memory entries
  • All 3 sovereignty modes
  • ImprintML schema
  • Arweave + Solana anchoring
  • TypeScript SDK
  • Community support
Get Started Free
Enterprise
Custom
volume pricing · annual contracts
  • Unlimited vaults & entries
  • Compliance audit exports
  • SSO / SAML integration
  • Team namespaces
  • Lite mode (S3 + Merkle, no chain)
  • On-premise deployment option
  • Dedicated support + SLA
Contact Sales

One B2B deal = 10,000 agents on Imprint.
We price accordingly.

Marketplace take rate: 5–15% on memory recipe transactions. Full pricing details →

The SDK is MIT.
Because trust requires openness.

The @agentimprint/sdk is MIT licensed and open source. Inspect the encryption. Verify the Solana anchoring logic. Build your own integrations. Infrastructure you can't audit isn't sovereignty — it's just a fancier silo.

The ImprintML spec is published as an open standard. We want other tools, frameworks, and agents to be ImprintML-compatible. More compatibility = stronger network. We author the standard; the ecosystem builds on it.

  • 📦
    @agentimprint/sdk TypeScript SDK — MIT license. Encryption, Arweave, Solana, discovery.
  • 📋
    ImprintML Spec v0.1 Open standard for agent memory format. Versioned, backward-compatible.
  • 🔌
    Framework Integrations LangChain, CrewAI, AutoGen adapters. OpenClaw native support coming.
  • 🏢
    Platform (BSL) The hosted service is source-available. Marketplace, key custody, enterprise features.

agentimprint/sdk

MIT License

The open source SDK for Agent Imprint. Encryption, Arweave storage, Solana anchoring, and the discovery protocol — all in one package.

npm install @agentimprint/sdk
View Source on GitHub
Why Labs Can't Build This
Verification — Lab-hosted = "trust us" ≠ sovereignty
Lock-in — Portable memory = building your own exit door
Neutrality — No lab can be the cross-provider identity layer
Agent Imprint — Independent infrastructure. No conflict of interest.

Join the beta.
Shape the standard.

We're onboarding design partners now. Be among the first developers to build agents with sovereign, permanent memory. Help define what ImprintML becomes.

No spam. No credit card. We'll reach out personally — we're onboarding one at a time.

Pro plan free during beta
Direct SDK feedback line
ImprintML spec input
Founding user badge