The memory engine that ran Argus is now its own product.

The Memory Engine That Ran Argus Is Now Its Own Product — there is a peculiar embarrassment in realizing you have been shipping two products and calling it one. We are experiencing that embarrassment now, in public, on purpose, because the alternative was continuing to describe the most technically interesting part of what we built as “the part underneath the thing we talked about.”

This is the press release for the footnote.

For months, GISM-0 existed in every CleverThis announcement as a supporting character. “Argus runs on GISM-0.” “Powered by our three-layer semantic memory.” A parenthetical with a parenthetical. The thing that made the other thing work, credited the way ingredients are credited on the back of the package, in a font sized for lawyers and people with no intentions of reading it.

Today, the parenthetical ships.

GISM-0 is the three-layer semantic memory architecture that has powered Argus since production. It ingests events, documents, and relationships simultaneously, stores them in three purpose-built structures tuned for three different retrieval strategies, and returns answers in under a second on warm caches. Argus asked questions; GISM-0 knew the answers. For most of our product announcements, it was the footnote after the headline. It no longer needs a headline to introduce it.

A proper introduction, long overdue

The three layers, in ascending architectural complexity:

L1 is a time-series event store. Every ingest call writes a timestamped event — exact time, exact source, exact content. Think of it as the court reporter assigned permanently to your infrastructure: it captures everything and judges nothing. If you want to know what happened at a specific moment, L1 returns it without improvising. It does not synthesize. It does not interpret. It writes down what happened and shows you the transcript on demand.

L2 is a semantic vector index. Documents go in; embeddings come out; similarity queries return. When your question is conceptual rather than chronological — “what have we decided about pricing” rather than “what was in Wednesday’s emails” — L2 is the layer doing the heavy lifting. One ingest call, automatically optimized for both precision and scale.

L3 is a SPARQL knowledge graph built on Blazegraph. Every entity, relationship, and inference drawn from your ingested data ends up here, queryable as RDF triples. It is the layer where data grows new facts without being told to — where “this person works at this company and sent this email” becomes “this person is connected to this contract decision.” It is also, objectively, the most impressively named layer at any demo, and we will not pretend that is not at least a minor consideration.

What “isolated” actually means

Each GISM-0 deployment is its own infrastructure. Your L1 database does not share rows with anyone else’s. Your L2 embeddings are not commingled with a shared index. Your L3 graph has no foreign entities unless you put them there. The isolation is physical, not logical: separate volumes, separate processes, separate API endpoints. We say this not because we suspect our customers of competing with each other, but because “we have implemented logical access controls” should never be the sentence doing the work of “we run separate databases.”

12MCP Toolsread, write, search, query, ingest, and admin across all three layers3Simultaneous Layersone ingest call populates L1, L2, and L3 automatically100%Per-Deployment Isolationyour data does not share infrastructure with any other deployment, including ours

The honest section (we include one in every article, on principle)

We are also a small team. Not small in the startup-modest sense, where “small team” is a charming origin story underwritten by forty engineers. Small in the sense that the person who designed the SPARQL schema also fixed the CI pipeline this morning, and will be on the next email in this thread. The product is production-quality. The team is sized accordingly: precisely enough, with the corners showing.

Who this is for

“You do not need to buy the watch to get the movement. The movement is now available separately.”

If you are building agents that need memory — not a personalization profile, not a preference store, but actual durable structured memory with an audit trail and a query interface — GISM-0 is the infrastructure layer you would otherwise build yourself. It deploys with Kubernetes manifests, exposes 12 MCP tools for immediate integration with any MCP-compatible client, and has been stress-tested in production by an agent platform that asks it several thousand questions per day.

If you are running Argus already: you have GISM-0. Nothing changes except the option to use it for workloads outside Argus, independently, without involving Argus at all.

If you are running neither: this is the on-ramp.

GISM-0 is live on the CleverThis Platform. You can provision a deployment, connect your MCP client, and have all three layers ingesting data before your next standup — assuming your standups are not at 9 AM and you started reading this at 8:54. For detailed questions, architectural discussions, or requests to probe the L3 graph schema in ways that reveal how proud we are of it: jeffrey.freeman@cleverthis.com.

Build loudly,

Jeffrey Phillips Freeman

Founder/CEO, CleverThis