A StricklySoft platform

Give your AI a memory it remembers — and a reason to stop guessing.

Lumen is a governed, self-hosted sovereign memory layer that grounds LLM agents in your organization's real, recorded knowledge — and structurally reduces hallucination.

Sovereign Memory for AI

How a governed, self-hosted memory layer makes LLM agents reliable.

The problem

Brilliant. But forgetful — and prone to invention.

Stateless

Every session starts from zero. Yesterday's knowledge is gone.

Hallucination

Missing a fact, the model fills the gap with a confident guess.

The answer

A memory your organization owns.

Lumen is a portable, governed memory layer — self-hosted on your own infrastructure. Nothing is handed to a model vendor to keep.

Polyglot persistence

The right store for every kind of memory.

Relational · system of record Document · full text Graph · relationships Vector · meaning Cache · speed Object · raw bytes
Retrieval by fusion

Search by keyword, meaning, and relationship — at once.

Query keyword · meaning · relationship Best match rises

The model answers from your recorded reality — not from a single brittle lookup.

Wired into the runtime

Memory that works without changing the model.

Session start Every prompt · ambient recall After each turn · saved Session end · summary
Why it's trustworthy

Five ways a governed memory cuts hallucination.

1 Grounding in real, retrieved data
2 A standing rule against fabrication
3 Verify-before-assert enforcement
4 Volatility flags on living facts
5 Continuous correction from its own errors

Memory as the product — not an afterthought.

Private.  Portable.  Foundational.

stricklysoft.com

Intro
The frontier problem

The barrier isn't capability. It's reliability.

Every standard model interaction starts from zero — and when a fact is missing, the default is a fluent, plausible guess. Statelessness and hallucination are the central barriers to trusting AI with real operational work.

Statelessness

Knowledge built in one session is gone in the next. The model never accumulates an understanding of your business, systems, or history.

Hallucination

The most common trigger for a fabricated answer is a missing fact. Lacking it, the model invents something confident — and possibly wrong.

How Lumen works

A sovereign memory layer, wired into the agent's runtime.

Lumen gives an agent a real, lasting memory it can read and write — and it follows the agent across tools, projects, and even different underlying models.

Polyglot persistence

No single database is good at everything — so Lumen uses specialized stores: relational (system of record), document, graph, vector, cache, and object.

Retrieval by fusion

Lumen queries multiple stores at once — by keyword, by meaning, and by relationship — then fuses the results so the best match rises to the top.

Event hooks

Session start loads standing rules + recent context; every prompt triggers ambient recall; every turn is saved; session end writes a durable summary — no model changes required.

The platform

A self-hosted AI substrate — everything in service of memory.

Lumen runs on a private, GitOps-managed Kubernetes platform whose entire configuration lives in version control, so the system can rebuild and update itself from a single source of truth.

Infrastructure-as-code

Provisions the machines, assembles the cluster, and continuously deploys every service — all driven from declarative source.

Shared SDK

A common library standardizing resilient DB connections (retry, circuit breaking, health checks), identity propagation, tracing, and lifecycle.

Nexus

The API gateway and orchestrator — the controlled front door managing sessions, quotas, and routing.

Vigil

An execution-control plane that governs how AI agents run tasks, keeping their actions inside policy.

Lumen

The memory layer — and the product the rest of the platform exists to serve.

Sovereign by design

You own the data and the infrastructure it lives on. Nothing is handed to a third-party model vendor to retain.

Reliability

Five mechanisms that reduce hallucination.

A governed memory layer removes the single biggest cause of fabrication — a missing fact — and structurally pushes the model to check rather than guess.

1 · Grounding in real data

Relevant facts are retrieved into context before the model answers — replacing "guess from training" with "answer from retrieved reality."

2 · A rule against fabrication

Every session: if the answer isn't in memory, say so — don't invent it. "I don't know" becomes a safe behavior, not a trigger.

3 · Verify-before-assert

At the moment the agent is most likely to guess, it's reminded to consult the source first: authoritative document → graph → guess (last resort).

4 · Volatility flags

Memories describing live, mutable state are tagged "verify before asserting" — preventing confident statements of outdated information.

5 · Continuous correction

When the agent gets something wrong, the fix is captured as a durable rule that fires later — the same mistake gets structurally harder to repeat.

Honest about limits

It reduces — not eliminates — hallucination, and only grounds the model on what's been captured. A system built to curb confident-but-wrong answers shouldn't make them about itself.

Why sovereignty is the point

Durable, owned, governed memory.

Private

The data never leaves infrastructure you own and control. No external vendor holds the institutional memory of your business.

Portable

The memory follows the work across tools, machines, and even different underlying models. Not locked to one vendor's ecosystem.

Foundational

It solves the one thing every model is bad at: remembering, accurately, over time.

PrivatePortableFoundational

Turn a capable-but-stateless model into a dependable collaborator.

See how Lumen and the StricklySoft platform give your AI a memory you own — and a reason to stay inside the truth.

Request a briefing →