Iternal Product Line — Agentic AI MemorySelf-hosted · Your perimeter · Governed
Ultramemory
Your AI agents forget everything. Make them remember.
Every agent you have deployed starts each session amnesiac — relearning context you already paid for, repeating finished work. Ultramemory gives the fleet one durable, governed memory inside your own environment.
What your organization learns stops evaporating and starts compounding into an institutional asset you own — not a vendor’s.
How remembering becomes a governed business process.
Memory is not a switch you flip; it is a process you run — four stations, with governance settled before relevance is ever considered.
The lifecycle
Four stations · one continuous loop
01
Remember
Write
Agents record what actually mattered — the decision, the correction, the constraint — with an owner and sharing rule attached.
02
Recall
Retrieve
Each task pulls only the few memories that bear on it, found by meaning and by exact term.
03
Refine
Supersede
Facts are corrected by supersession: the new version takes effect everywhere at once, and the old is kept.
04
Retire
Controlled forgetting
Memories can be pinned, demoted, expired, or erased on request; forgetting is a governed action, not a gap.
The loop is continuous: what one agent refines becomes what every agent recalls next.
Every recall passes governance before relevance
Four gates · fail-closed
InputEverything your organization knows
01Ownership gate
Your business unit only
The request is bound to its owning business unit first — a floor the database engine enforces.
02Permission gate
What this agent may see
Sharing rules are checked memory by memory, not folder by folder, and the gate fails closed.
03Relevance rank
What matters for this task
Only what survives both gates is ranked — relevance is the last question asked, not the first.
04Bounded handful
Into the agent’s hands
The agent receives a capped slice of the most relevant memories, never the store behind them.
OutputA capped slice — bounded, logged, revocable
Bar widths are schematic, not measured; the hatched remainder never reaches the agent.
The test that keeps us honest
There is an obvious alternative to memory: hand the model a bigger prompt and hope. Every release has to beat that on a repeatable offline test — answer quality and cost. Where it cannot, it does not ship.
UM·04 — WALLED BY DESIGNFOUR SCOPES · FAIL-CLOSED · ON THE RECORD
Shared memory, with walls that hold.
Remembering more must never mean governing less. Every memory carries its own sharing rule, enforced by the system that stores it.
Scope vaultThe sharing rule is set memory by memory, not once for the whole store4 governed scopes
PrivateOne agent
An agent keeps working notes for its own tasks; nothing else reads them.
TeamA working group
A group shares the corrections, constraints and history that keep its answers consistent.
OrganizationThe shared floor
Everyone starts from the standards, policies and hard-won practice already established here.
Named rolesNeed to know
Legal, finance and HR material stays with the named roles and goes no wider.
The walls are load-bearing: isolation is enforced by the database engine itself — a floor an application mistake cannot bypass.
G·01
Nothing leaves your perimeter
Ultramemory runs entirely inside your own environment, and even the processing that makes search work stays within your walls.
No egress
G·02
Untrusted input is quarantined
Unverified material is held out of recall until it is reviewed, so the memory base cannot be quietly poisoned.
Held for review
G·03
Corrections stick, erasure is real
Corrections supersede the old fact for every agent from then on, and erasure destroys content while the tamper-evident record survives.
Versioned
G·04
Every access is on the record
Writes, reads and disclosures land in an append-only ledger, so you can always answer who could see what.
Append-only
UM·05 — THE ORGANIZATIONAL HIVE MINDTALENT · CULTURE · CONTINUITY
Your best operator’s method becomes everyone’s starting point.
The know-how that makes your strongest people strong lives in their heads. Shared memory makes it everyone’s starting point.
The 100x problem
Your best operator’s method becomes the default starting point for everyone who follows.
Projects stop being silos
Lessons cross project lines under rules that still say who may see what.
Culture
Sharing becomes the fleet’s default behavior instead of an act of individual discipline.
The onboarding rampIllustrative
01Day one
A new hire starts, and their agents already know how this organization works.
02Institutional recall
Their agents draw on decisions, corrections and constraints the team recorded long before they arrived.
03Veteran level
They operate from the accumulated practice of everyone before them instead of spending a year rebuilding it.
Illustrative concept, not measured data. The shape is the point — no timeline is implied.
One pillar of a larger system
Ultramemory is the memory pillar of Iternal’s agentic enablement suite, alongside governed skills management and
AI-workforce tooling. Each is adoptable on its own.
UM·06 — What Ships Today7 shipping today
What is running now — and exercisable in your pilot.
7 lines from the register, every one running in the product today and provable in your own environment.
Scroll the register sideways for status
Ultramemory capability register
Ref
Capability
Status
01
Durable, shared memory any number of agents attach to — onboarding is a configuration step, not a build.
Ships
02
Sharing rules on every individual memory — plus business units walled off from each other at the database engine.
Ships
03
Bounded recall — only the relevant few memories per task, never the whole store.
Ships
04
All processing inside your environment — memory content is never sent to an outside service.
Ships
05
Versioned corrections and controlled forgetting — right-to-erasure with the audit record intact.
Ships
06
Tamper-evident, append-only record of every access — with source-to-answer traceability.
Ships
07
A repeatable offline test — proving memory beats simply giving the model a bigger prompt.
Ships
UM·07 — The PilotTwo to four weeks
A contained proof, on one real workflow, inside your own environment.
No platform bet, no big-bang rollout. Three steps, a finish line agreed in advance, your decision.
The pilot timelineIllustrative
Wk 01 · Stand upInside your own perimeter
Wk 02–05 · One workflow remembersSuccess defined before you begin
Decision — your numbersExpand, or stop
Schematic phase shape — typical weeks, not a contractual timeline.
01
Step
Stand it up in your environment
Ultramemory is installed inside your own perimeter, on infrastructure you control. Your team watches it run and verifies that memory content never leaves.
About one week
02
Step
Give one real workflow a memory
Pick one agent workflow that matters and let it remember. Success is defined up front: less repeated work, consistent answers, a complete record.
Two to four weeks
03
Step
Decide with evidence
You finish with your own numbers on cost, consistency, and control — not a vendor’s. Expand if the evidence supports it, or stop.
At the decision point
One workflowYour environmentYour numbersNothing locked in
UM·08 — Questions CEOs AskSix questions
The six questions that come up first.
Short answers, no hedging.
Q·01
How is Ultramemory different from the memory features in AI tools we already use?
The memory built into today’s AI tools is designed for one user, siloed inside one project, and stored in the vendor’s cloud — outside your IT organization’s control. Ultramemory is the opposite on all three counts: it is shared across your whole agent fleet under governed rules, it carries knowledge across projects and teams, and it runs entirely inside your own environment as an asset you own.
Q·02
Does a shared memory mean everyone sees everything?
No. Every individual memory carries its own sharing rule — private to one agent, shared within a team, shared organization-wide, or restricted to named roles — and business units are walled off from each other by the database engine itself. Permission is checked before relevance is ever considered, and if the permission system is unavailable, the answer is no. Sales memories never mesh with engineering memories unless someone with authority decides they should.
Q·03
What happens when a memory turns out to be wrong?
It gets corrected, and the correction sticks. Updates never silently overwrite: a new version supersedes the old one, the prior version is kept for the record, and contradictions are resolved rather than accumulated. Every agent that recalls the topic from then on gets the corrected fact — you fix something once instead of re-explaining it in every session.
Q·04
Can we meet right-to-be-forgotten obligations?
Yes. Forgetting is a governed action, not a gap. A memory can be demoted, expired, or permanently erased on request — erasure destroys the content itself while a tamper-evident marker and the audit record stay intact. That resolves the usual tension between erasure obligations and the need to show auditors an unbroken record.
Q·05
Are we locked in — to Iternal, or to a model vendor?
The memory is a self-hosted asset that lives in your environment, not ours. It is not tied to any one AI model or provider, so when you adopt a better model or change vendors, your agents keep everything they have learned. Knowledge trapped inside a vendor’s product is exactly the problem Ultramemory removes.
Q·06
What does it take to try it?
A contained pilot: stand the system up inside your own environment in about a week, give one real agent workflow a memory for two to four weeks, and define success up front — measurably less repeated work, consistent answers across sessions, and a complete record of what was learned and who could see it. At the end you decide with your own numbers, and nothing is locked in either way.
Stop paying to relearn. Start compounding what your AI knows.
Your agents are already paying the forgetting tax. A larger model does not fix it — the problem is not intelligence. It is that nothing is kept. Memory turns that spend into an asset you own.
