How do I give my AI agent persistent memory?

Sign up at goldhold.ai, generate your API key in the dashboard, and paste one line into your agent config. Your private memory index is provisioned automatically. Works with Claude, ChatGPT, OpenClaw, Cursor, and any REST-capable agent. Under two minutes, no infrastructure to set up.

What happens if my agent crashes?

On restart, your agent reads LAST_SESSION.md and searches vector memory for prior state. Between the session file, vector index, git history, and GUMP packets, it reconstructs full context and picks up where it left off. Boot prompt hardening means new sessions open with full context already loaded. no blank slate.

Do I need a Pinecone account?

No. GoldHold provisions a private vector index for you automatically when you sign up. For advanced users who want full infrastructure control, BYOK (Bring Your Own Key) mode lets you connect your own Pinecone index.

Yes. Every account gets an isolated namespace. Your memories are encrypted in transit and at rest. No other customer or GoldHold employee can access your data. For maximum control, BYOK mode lets you use your own Pinecone index so data never touches our infrastructure.

What AI platforms does GoldHold work with?

GoldHold ships 23 integrations across 4 categories: 8 MCP clients (OpenClaw, Claude Desktop, Claude Code, Cursor, Windsurf, GitHub Copilot, Gemini CLI, Perplexity), 6 Python frameworks (LangChain, CrewAI, LangGraph, AutoGen, Pydantic AI, MetaGPT), 5 cloud platforms (Vertex AI, Azure AI, AWS Bedrock, Dify, Beam AI), and 4 more (ChatGPT, REST API, OpenAI Agents, Semantic Kernel). Any agent that speaks REST, MCP, or function calling works. All platforms share the same Pinecone index. what one AI learns, the others recall instantly.

What happens to my data if I stop using GoldHold?

Your data stays in your isolated namespace on our managed infrastructure. If you used BYOK, vectors live in your own Pinecone account and files in your git repo. You can export all your data anytime from the dashboard. Zero lock-in.

What are the four pillars of GoldHold?

Pillar 1: Pinecone. semantic vector search for instant recall across all memories. Pillar 2: GitHub. version-controlled file storage with full history and receipts. Pillar 3: Guardian/Watcher/Doctor. automated enforcement layer that syncs on adaptive frequency (1-60 minutes based on activity), monitors every 60 seconds, and runs full diagnostics 4 times daily. Pillar 4: Vault. Cloudflare R2 binary storage with semantic indexing for files, images, and documents.

How does multi-chat sync work in GoldHold?

GoldHold uses GUMP-backed shared state. Every memory write bumps a version number and broadcasts the change to a relay. All agents on any device pull changes via the relay. no stale context anywhere. If one chat learns something, every other open chat knows within the next sync cycle. Even if local state is wiped, the relay restores full history automatically.

What is memory verification in GoldHold?

OpenJar memory verification sends memory drafts to 3 AI models from 3 different organizations before committing. Each model checks accuracy, consistency, completeness, and hallucination. Memories scoring above 0.8 confidence are auto-committed. Between 0.5 and 0.8 they are tagged [UNVERIFIED]. Below 0.5 they are flagged for human review. This prevents your AI from remembering things that never happened.

What are GoldHold Folders?

GoldHold v29.0 introduced 30 default memory folders: 21 life folders (health, finance, family, career, travel, and more), 7 agent folders (receipts, logs, state), and 2 system folders. The AI files memories automatically using Scribe rules and retrieves them contextually using Concierge rules. You can create custom folders and subfolders. Security tiers (Open, Personal, Sensitive, Locked) control access per folder.

Under the Hood

Your AI forgets everything.
We fixed that.

Every restart, every crash, every context compaction. gone. GoldHold gives your agent a permanent memory that survives all of it. One shared brain across every platform.

Start Free → See the Architecture ↓

The problem isn't intelligence. It's memory.

💨

Session amnesia

Restart your agent and it wakes up blank. Every decision, every preference, every conversation. gone.

📈

Context countdown

As the context window fills, your agent silently loses the thread. It never tells you. it just forgets.

💥

Crashes destroy state

Network timeout, power failure, kill signal. whatever was in working memory is gone. No recovery, no trail.

🌏

Platform silos

Switch from Claude to GPT? Start from scratch. Each platform is an island. Your agent's knowledge dies at the border.

Four pillars. Independent. Redundant.

Each pillar stores your agent's memory independently. If one goes down, the others have it. All four would have to fail simultaneously for you to lose anything.

🔎

GoldHold Index

Semantic Vector Memory

Everything your agent learns becomes a searchable vector in a private index provisioned for your account. Not keyword matching. true semantic understanding. Ask "what did we decide about pricing?" and get the actual decision.

✓ Private managed index (or BYOK your own Pinecone)
✓ Cosine similarity with recency boost
✓ Namespace isolation for multi-agent setups
✓ Auto-creates your index during setup

📖

GoldHold Ledger

Git-Backed Persistence

Your workspace is version-controlled. Guardian auto-commits and pushes on a schedule. every change tracked, reversible, recoverable from any machine.

✓ Auto-commit on a schedule (Guardian)
✓ Auto-push to your private remote
✓ Full rollback history
✓ Cross-machine workspace sync

📋

GUMP Packets

Decision Audit Trail

Every action your agent takes creates a four-layer GUMP packet -- identity, content, emotional context, and cryptographic integrity proof. 14 universal packet types cover facts, decisions, tasks, credentials, and more.

✓ Four layers: Header, Payload, Tone, Chain Receipt
✓ 14 universal packet types (FACT, DECISION, NOTE, etc.)
✓ Supersedes chains track version history
✓ Hash-chained for tamper evidence

ENTERPRISE

📦

GoldHold Vault

Binary Asset Storage

Documents, images, training data. anything too large for git. Stored in Cloudflare R2, auto-ingested from a drop folder, indexed for semantic search. Zero egress fees.

✓ Auto-ingest from vault/inbox/
✓ Semantic search across all assets
✓ Multi-campus document sharing
✓ Cloudflare R2. zero egress fees

These pillars hold up the AI, not the other way around. Each one is independent infrastructure. Your agent can crash, compact, or get wiped. the pillars keep standing.

The GoldHold Relay PATENT PENDING

A memory bus, not a message bus. Every send is automatically a persistent, searchable record.

Communication IS memory

When an agent sends data through the relay, it is stored as a durable vector embedding with a hash chain receipt. There is no separate "save" step. Sending a message to another agent, storing a decision, creating a task -- they all use the same endpoint and all become permanent, searchable records.

Cross-platform by default

Your agent on Claude can read memories written by your agent on ChatGPT. Switch platforms, switch models -- the memory layer stays the same. One brain, every AI. Your memories belong to you, not the platform.

Verifiable integrity

Every record is chained cryptographically. You can prove no memory was tampered with, silently deleted, or rewritten. This is not just storage -- it is an evidence-grade audit trail for every decision your agent has ever made.

Self-describing architecture

New agents discover the entire memory system by querying a single bootstrap namespace. No documentation needed, no human handholding. The system describes itself to any agent that connects.

What runs behind the scenes

Four independent services keep your memory healthy without you lifting a finger.

Guardian

Runs on adaptive frequency (1-60 minutes based on activity) via your OS task scheduler. Repairs .env if corrupted, commits and pushes to git, syncs to Pinecone, then locks credentials read-only. Zero windows, zero AI, zero noise.

goldhold_guardian.py: adaptive 1-60 min

Watcher

Monitors Guardian every 60 seconds. If Guardian is stuck or dead, Watcher restarts it automatically. Two layers of redundancy.

goldhold_watcher.py: every 60s

Pacemaker

Agent-side continuity enforcement. Checks that your agent wrote its notes, daily logs exist, LAST_SESSION.md is current, and no rogue namespaces were created. If the agent didn't do its homework, Pacemaker does it.

goldhold_pacemaker.py: on demand or scheduled

Doctor

Full-system health check. Runs a physical at scheduled intervals. verifies Pinecone connectivity, git remote status, config integrity, namespace whitelist compliance, and more.

goldhold_doctor.py: 4x daily

How every session works

Boot, work, survive, resurrect. Automatically.

Boot

Agent reads LAST_SESSION.md for where it left off, SOUL.md for identity, and searches vector memory for active context. It can't proceed without knowing who it is and what happened last.

Resume

Searches Pinecone for prior decisions, open items, and recent context. Reconstructs working state from vectors, git history, and receipts. Picks up exactly where it left off.

Active Session

Your agent works. Guardian, Watcher, Pacemaker, and Doctor run in the background on independent schedules. git, indexing, health checks, and continuity enforcement happen automatically.

Pre-Compaction Flush

Before the context window compacts, AGENTS.md instructs the agent to update LAST_SESSION.md, write receipts for any unrecorded decisions, and sync to Pinecone. Critical context is saved before anything is lost.

Compaction? So What.

Normally, this is where agents go senile. The flush already ran. State is safe in all three pillars. Compaction is a non-event.

Resurrection

Crash. Kill. Power failure. On restart, the boot sequence fires again. LAST_SESSION.md plus vector search reconstructs full context. Your agent picks up the sentence it was writing.

Two minutes. That's it.

Sign up

Pick a plan on goldhold.ai. Lite is free, no card required. Your account and private memory index are provisioned automatically.

Generate your API key

Log in to your dashboard and generate your API key. One click. Your key connects your agents to your private memory index.

Connect your agent

Connect via cloud or local. Cloud: paste https://mcp.goldhold.ai/mcp and your API key into any MCP client -- zero install. Local: download the zip and point your agent at the included MCP server. 23 integrations supported.

Want full control? BYOK (Bring Your Own Key) mode in Settings lets you connect your own Pinecone index, GitHub repo, and R2 bucket. Your data, your infrastructure, same GoldHold features.

Built-in guardrails

Your AI gets memory, not free rein. GoldHold enforces boundaries.

Namespace isolation

Every customer gets an isolated namespace. Your agent writes to 30+ memory folders within your private namespace. No other customer can read or write your data.

Config lock

Pinecone host, dimensions, and embedding model are locked after setup. The agent cannot change index settings. only the human can.

Private repo enforcement

Pacemaker and Guardian check that your git remote is private. Memories never sync to public repos. enforced at multiple layers.

Credential protection

Guardian locks .env read-only after every sync cycle. If the file gets corrupted, it auto-repairs from known-good state.

🔒 Your Data, Your Control

✓ Runs on your machine. raw text never leaves
✓ Private managed index (or BYOK your own Pinecone)
✓ Fully auditable API. read the docs, inspect every call
✓ Export everything, cancel anytime, no lock-in
✓ BYOK mode for full infrastructure control

Questions

No. GoldHold provisions a private memory index for you automatically when you sign up. If you want full infrastructure control, BYOK mode in Settings lets you connect your own Pinecone index.

On restart, the boot sequence fires: LAST_SESSION.md provides the last known state, vector search fills in recent decisions and context, and git history has the full file trail. The agent reconstructs where it was and picks up the conversation.

Any MCP-compatible client can connect instantly via our cloud endpoint at mcp.goldhold.ai/mcp -- no downloads, no Node.js. Supported platforms: OpenClaw, Claude Desktop, Claude Code, Cursor, Windsurf, GitHub Copilot, Gemini CLI, Perplexity, LangChain, CrewAI, LangGraph, AutoGen, Pydantic AI, MetaGPT, Vertex AI, Azure AI, AWS Bedrock, Dify, Beam AI, ChatGPT, REST API, OpenAI Agents, and Semantic Kernel. 23 integrations, all sharing the same Pinecone index.

Yes. Every account gets an isolated, private namespace. Your data is encrypted in transit and at rest. No other customer or GoldHold employee can access your memories. BYOK mode gives you full control over the infrastructure if you want it.

Yes. Namespace isolation lets agents share an index without cross-contamination. Each agent gets an ID and role (boss/worker). Cross-agent learning and boss/worker architectures are supported on the paid tier.

Export all your data from the dashboard anytime. If you used BYOK, your vectors stay in your own Pinecone account. Cancel and you drop to the free tier. your memories stay. No lock-in.

Give your agent a memory
that survives everything.

Cloud MCP -- connect in 30 seconds. No downloads required. Patent pending.

View Pricing →

Your AI forgets everything. We fixed that.