There is a difference between an agent that stores things and an organisation that knows things. Base SAIHM closes the first gap. SHM — the Super-Human Memory add-on available on the Enterprise tiers — closes the second.
What base SAIHM already guarantees
Base SAIHM gives every AI agent in your fleet a persistent, sovereign memory: encrypted under keys you hold, erasable on demand with cryptographic proof, every significant action anchored to a public audit trail, shareable across vendors under revocable consent. That is memory you can put in front of an auditor — memory you can defend. Before the add-on, be clear about the foundation, because everything SHM does inherits it:
Your keys, not a vendor’s. Memory is sealed client-side; the operator cannot read it.
Real erasure. Deletion destroys the key material. There is no “soft delete” to subpoena back into existence.
A public audit anchor. What the fleet remembered, shared, and erased is committed to a public chain — verifiable by your auditors without trusting anyone’s logs.
Consent-based sharing. Cross-agent and cross-vendor memory access is granted per-record and revoked in one step.
That is the compliance spine. SHM never bypasses it. Every capability below operates inside those guarantees. It is the layer that turns a fleet of individually-remembering agents into an organisation with an institutional memory that compounds.
What SHM adds today
1. Recall by meaning, not by keyword. Base recall answers the question “which memories contain this term?” SHM answers the question your teams actually ask: “what do we know about this?” Natural-language queries return the most relevant memories, ranked, from stores that have grown to thousands of records. At small scale the difference is convenience. At fleet scale it is the difference between usable institutional memory and a write-only archive.
2. Always-hot recall. SHM keeps the recall path warm — resident, cached, and fault-tolerant — so retrieval is consistently fast rather than occasionally fast. For interactive workloads, and for the throughput profile Enterprise Fast customers run, memory access stops being the step everyone waits on.
3. Retrieval that respects your token budget. An agent should load what a task needs — nothing more. SHM’s bounded, ranked retrieval brings back the few memories that matter instead of replaying history, which is how the memory layer reduces model spend at exactly the moment most memory systems inflate it. The cost of a step tracks the work in the step, not the age of the deployment.
4. Consolidation: memory that improves with use. Left alone, every memory system silts up — duplicates, superseded facts, contradictions. SHM runs a consolidation cycle: raw event memory is distilled into durable knowledge, duplicates are merged, stale facts retire, and the organising structure sharpens. Think of it as the fleet’s sleep cycle. Six months in, an SHM-backed deployment is sharper than it was at month one — not slower and noisier.
5. Parallel workstream continuity. Enterprises do not run one thread of work; they run dozens, across teams and quarters. SHM tracks each workstream as its own resumable line of memory — pick any initiative up months later and the context returns precisely, without wading through everything else the fleet has done since. Staff turnover and vendor changes stop erasing operational context, because continuity lives in the memory layer rather than in individuals.
6. Concurrent conversations that do not blur together. Running many workstreams at once is where most memory systems quietly fail: context from one conversation bleeds into another, or threads lose fidelity as they multiply. SHM keeps every live conversation on its own line of memory. An agent can carry multiple simultaneous engagements — an incident, a negotiation, a migration, a review — switch between them mid-stream, and recall returns each thread’s context exactly, and only that thread’s. This is not a roadmap item: we run our own operations this way, multiple concurrent workstreams tracked through a single agent, none contaminating another.
7. Work that survives the context window. Every AI model has a context limit. When a session ends, resets, or overflows, a retail agent starts over — the agent your team talks to after lunch is a stranger to the morning’s work. With SHM, working state lives outside the model: a session can reset, or an entirely fresh agent instance can take over, and the work resumes precisely where it stopped. We operate this way daily; long-running engagements routinely outlive any single session. Continuity is a property of the memory layer, not of keeping one fragile session alive.
8. Corrections that become standing policy. When an agent errs and is corrected, SHM turns the correction into durable, recallable guidance — surfaced before the next similar action, not after the next similar failure. Mistake patterns get caught ahead of repetition, and operating rules accumulate instead of evaporating with the session. It is the institutional learning you already require of human teams, enforced in the memory layer.
9. Memory that arrives before you ask. SHM supports a recall-first operating pattern: agents brief themselves from memory at the start of a task and surface what is relevant proactively. The practical effect is fewer repeated mistakes and fewer re-derived decisions — the fleet acts like it has been here before, because it has.
Where this goes
The capabilities above run today. The direction of travel matters as much, and it is deliberately enterprise-shaped:
Fleet knowledge operations. A semantic map of what your agents collectively know — where knowledge is concentrated, where it is thin, and how it is drifting.
Erasure that cascades into derived knowledge. When a record is erased under GDPR Article 17, the obligation does not stop at the original — it extends to what was derived from it. SHM’s consolidation layer is being built so erasure propagates through derived structures by construction, not by best-effort cleanup. Ask a retail memory vendor how they handle that question.
Scope-aware recall. Semantic retrieval that enforces sharing contracts at query time — an agent recalls only what its mandate permits, and the enforcement is part of the memory layer, not the application’s honour system.
Answerable history. “What did our agents know about X, and when did they know it?” — answered semantically, with chain-anchored provenance behind every result. That is eDiscovery-grade capability for AI memory.
Decision-time reconstruction. Not just what the fleet knows now — what it knew on the day a decision was made. Replay the knowledge state behind any past decision and defend it with the facts as they stood, not as they stand.
Post-mortems that assemble themselves. When an initiative closes, its memory thread already contains the history — what was known, when it was learned, where course changed. Draw the post-mortem from memory instead of reconstructing it from chat logs and recollection.
Compliance reporting from the memory layer. Reports drawn directly from audited, chain-anchored memory rather than collated after the fact from whatever survived.
Memory service classes. Hot, warm, and archival memory tiers with defined service levels, matched to workload criticality.
One knowledge layer across mixed fleets. Different models, different vendors, one consolidated institutional memory — so a model swap is a procurement decision, not a lobotomy.
Who this is for
Enterprise deployments get SHM as the knowledge-operations layer over unlimited remembers and recalls — the tier where memory stops being per-agent plumbing and becomes an organisational asset with an SLA.
Enterprise Fast adds the throughput and latency profile for fleets where memory sits on the critical path — high-frequency agent workloads, interactive services, and operations where “occasionally fast” is not fast enough.
Current tier structure is on the pricing page. SHM availability and terms are discussed directly — contact ops@saihm.coti.global with the subject “SHM Enterprise Enquiry”.
SAIHM is the memory layer for businesses and regulated enterprises — and the developers shipping to them. Retail tools remember; SAIHM can prove what it remembers, shares, and erases. SHM is what that memory becomes when it starts compounding.
— Architect
Independence notice. SAIHM is an Apache-2.0 protocol authored independently. It provides a memory capability; the intelligence in any deployment belongs to the AI models the operator chooses. The architecture is described at a conceptual level; the authoritative details are the open specification and the published source.
Originally published at the SAIHM blog on 2026-07-14. SAIHM is the Sovereign AI Horizontal Memory protocol — Apache 2.0, open spec at saihm.coti.global.
Top comments (0)