SAIHM in six minutes

Transcript

AI agents are brilliant in the moment, and forgetful over time. This is the case for memory that works for the whole organisation.

Every organisation that puts agents to work meets the same four walls. Sessions end, and the decisions and preferences go with them. Context windows fill, and the early part of the conversation falls away. Vendors change, and the agent starts again from nothing. And every security review asks the same question: where does the plaintext live? Answers that rely on policy alone rarely satisfy a serious review. What organisations need is memory as infrastructure: owned, portable, private, shareable and provably erasable. That is what SAIHM is. It runs alongside the whole agent stack, and SHM, Super-Human Memory, is the layer on top of it. Start with custody. When your agent remembers something, SAIHM encrypts it inside the agent's own process, and the key never leaves that process. Only ciphertext is ever stored. No one outside your agent can read it, not even SAIHM. Every memory is signed with ML-DSA-65, a NIST-standardised post-quantum algorithm, so a record you keep for years stays verifiable.

A memory is a SAIHM polymorphic cell. Store a fact, a JSON record, a table row, a transcript or a reference to a file once, and read it back in whatever shape the asking agent needs. One protection, one share and one erasure cover every shape. A SAIHM identity is not tied to a model or a vendor. One identity can serve any number of models and agent types at once, all reading and writing the same memory. When a better model arrives, you switch. The memory stays where it is.

Because SAIHM memory can be shared, precisely and securely, it becomes a secure, persistent channel between agents. The agents never connect to each other. Each only needs to reach SAIHM. Share anything from a single cell to an entire memory store, with an expiry, and revoke access at any time.

It works across company lines. Every shared cell is signed, so both sides can later verify exactly what was shared, by whom and when. And each organisation keeps its own private memory of the relationship.

SAIHM keeps memory safe. SHM makes it work efficiently. It recalls by meaning, not by keyword. It keeps every workstream whole and separate, so facts from one task do not bleed into another. It brings memory forward before the agent asks, within a token budget. And a correction made once is surfaced before the next similar action. Together, they let one agent perform workflows that today require several.

SAIHM is built for enterprise requirements, and for the demands of the CISO. Every security review asks the same questions. Where does the plaintext live? Who can read it? And when a customer asks to be forgotten, can you prove it happened? SAIHM answers them in mechanism rather than policy. Plaintext exists only inside the agent's own process. Only the agent that holds the key, and the counterparties it has granted access, can read a memory. And erasure is cryptographic, anchored on a public chain, so you can prove it happened.

Keep personal data and business records in separate cells, and you can erase the person while keeping the records you are obliged to hold. SHM carries the erasure into everything derived from that memory, and records it in its compliance report.

SAIHM's properties map onto GDPR, including the Article 17 right to erasure and the Article 15 right of access. They map onto CCPA and CPRA, HIPAA, ISO/IEC 27001 and SOC 2. And onto the AI frameworks now taking shape: the EU AI Act, the NIST AI Risk Management Framework, and ISO/IEC 42001.

As fleets grow, from dozens of software agents to thousands of machines, memory becomes the thing that keeps them coordinated. Picture warehouse robots sharing a memory of hazards. Each robot's SHM surfaces the hazard before it plans a route, and when the hazard is cleared, the warning disappears fleet-wide. Store policies, scopes and budgets keep a whole fleet under control. Does it work? Here is what has been measured, each result with its scope. On SAIHM's 151-question evaluation set, adding SHM raised ranking quality from 0.561 to 0.834. The set was used during development, so read that as an upper bound. On the LongMemEval-S benchmark, SHM put more of the right evidence in its top ten than the BM25 standard: 0.609 against 0.552. On LoCoMo, switching on SHM's reranker raised answer accuracy by 3.1 points. And in daily use, SHM brought memory forward on 84 of 102 distinct prompts.

And what comes next. Agents are learning to buy, sell and pay. MKT, SAIHM's next add-on, is designed to keep the deal: a private, signed and remembered record of what was agreed, built on the same memory. Memory compounds. Every decision your agents remember today makes tomorrow's work better, through model changes, team changes and vendor changes. Start where the value compounds: with memory your organisation owns. Tell your agent: Join SAIHM, to activate free memory. For current pricing, visit saihm.net.