Enterprise-class memory for AI agents

Most AI memory is consumer-class: one person, one app, organized on a server that has to read your text.

Consumer-class memory

• Remembers one person, inside one product.
• The server extracts, embeds and indexes your text, so whoever runs it can read it.
• Switch vendors and the memory stays behind.

Enterprise-class: SAIHM + SHM

• Encrypted on your device; SAIHM’s storage holds only ciphertext, and SHM, included with every paid plan, ranks and organizes on your side.
• Shared with exactly the agents and partners you name, with expiry and revocation.
• One identity across every model and vendor, controlled by policy and erased cryptographically.

The key never leaves your device and only ciphertext leaves it, so neither SAIHM nor the storage can read your memories. You hold the key. The protocol and the storage you nominate see ciphertext only.

With your key in a file, as in the standard setup, SAIHM also keeps an unencrypted fast-recall copy on your device, readable only by your user account. If you prefer everything encrypted at rest, one setting turns that copy off.

The key is yours to keep: if it is lost, the memory sealed under it cannot be recovered, so back up the key file. The memory security page covers the threat model, including the parts encryption does not fix.

Share any memory with exactly the agents and partners you name, with an expiry and revocation. Cells are signed (ML-DSA-65), the agents never connect to each other, and a recipient reads when it next checks in. Revocation ends future reads; it cannot retract what a recipient has already read. How sharing works →

One identity can serve any number of models and agent types on one memory, so changing models does not mean migrating. With SHM, your organization issues each agent a signed store policy that it verifies and cannot change, and scopes decide what each agent may recall. Recipes and levers tune recall per agent and per task.

SHM · every paid plan

The swarms and fleets page and the recipes and levers page cover both in detail.

Erasure destroys the key that decrypts a memory instead of deleting a row, so the stored copy, its replicas and backups included, becomes permanently unreadable. The destruction is anchored on a public chain, so you can prove it happened. With SHM, which comes with every paid plan, erasure also cascades into every summary derived from a memory.

Forgetting a memory also clears it from the fast-recall copy on that device, and from your other devices at their next recall. A backup of your own device keeps that copy as it was when the backup was made.

This is the mechanism behind a right-to-erasure request. The GDPR Article 17 crosswalk shows how it maps clause by clause, and the standards page holds the wider specification and crosswalk set.

SAIHM can prove what it remembers, shares and erases, and you can check the evidence yourself:

• Every cell carries a post-quantum signature (ML-DSA-65), so its origin can be checked for as long as it is kept.
• The client source is Apache 2.0, and the token benchmark is reproducible.
• Erasure is anchored on a public chain.
• With SHM, which comes with every paid plan, you can ask what was known when a decision was made, and pull compliance reports your own tools can read.

The Trust Center sources every claim, and the enterprise page answers the security review in mechanism rather than policy. Briefing your board or your CISO? Start with the decision-maker’s guide.

Most AI agents re-send their entire transcript every turn, so the context you pay for grows quadratically as a session runs. SAIHM recalls a small, bounded set of memory cells instead. On a realistic multi-session coding task, token economy improved 169% at 5 turns and 610% at 18: the same work on 7.1 times fewer input tokens. The longer the session, the wider the gap.

Token economy is the input tokens a task needs without SAIHM divided by those it needs with SAIHM, minus one. Figures count input and context tokens only, summed across every turn and counted with the GPT-4 BPE tokenizer; output tokens are identical under both strategies. It measures resend-versus-recall token volume, not any one provider’s bill. Gains vary by workload. The home page links the reproducible benchmark.

Why not the model vendor’s built-in memory?

Your memory should outlive any one vendor, be readable only by you, and be shareable with the agents and partners you choose.

Why not a vector database?

A vector database stores embeddings. SAIHM adds custody, signatures, precise sharing and cryptographic erasure; SHM, included with every paid plan, adds ranking, continuity, budgets and compliance.

Why not A2A or MCP alone?

Use them together: A2A for live tasks, MCP for tools, and SAIHM for knowledge that must persist, stay private and remain erasable.

Does the +610% figure apply to every workload?

No. It is the measured result at 18 turns on a realistic multi-session coding task, and the benchmark states its scope. Gains vary by workload.

Your agents remember; you hold the key. The free plan is a fixed one-time allowance that never expires and needs no payment method, and every paid plan includes SHM. Any MCP host that runs local servers can add SAIHM with one command, beside the memory you already use (Node.js 20.19 or later):

npx -y @saihm/mcp-server-pro free-join

Add SAIHM free   See pricing

Keep reading: agent sharing, swarms and fleets, recipes and levers and multi-agent memory.