An argument in five parts · then a product

Everyone has the same AI.
Nobody has your context.
Own it.

The models are becoming a commodity. What your AI knows about you — your business, your terms, your way of working — is not. Owning that context is the differentiator. Curating it, and retrieving only what each moment needs, is the force multiplier.

↓ five minutes · worth it
FACTS
119
WRITES / 7D
48 ▲269%
AUTO-MERGE RATE
90%
NEEDS REVIEW
2
What your agents learned today timeline →
Chargeback disputes now route through the EU disputes queue claude · customer-support · from 14 support conversations merged·auto
Escalation SLA is 4 business hours for tier-2 accounts jordan · customer-support · amended by hand merged
Refunds over $500 need director sign-off claude · contradicts f-1225 ($250) → held for your review held·review

a context graph at work — agents writing, policy deciding, one item waiting for a human

01 · WHAT JUST CHANGED

Intelligence stopped being the scarce ingredient.

You, your competitor, and a teenager with a laptop now call the same frontier models. When everyone reasons with the same brain, the output differs only by what goes into it: the accumulated, specific understanding of how your world works.

That understanding has a name — context — and right now, almost everyone treats it as exhaust. It gets typed into prompts, scattered across per-tool config files, and thrown away at the end of every session. The teams pulling ahead are the ones treating context as an asset: accumulated deliberately, owned outright, compounding over time.

the same frontier model + no context = everyone's answer
the same frontier model + your context graph = your answer
02 · THE THREE WAYS CONTEXT DIES

You're losing it, polluting it, or drowning in it.

Every team using AI today is running one of three failure modes — usually all three at once.

EVAPORATION

It doesn't survive the session, or the tool

The agent figures out how your deploy process works at 4pm. At 5pm the session ends and the knowledge is gone. Tomorrow it learns it again, on your token bill. Open a different tool and it never knew.
POLLUTION

Nobody curates what sticks

In March the agent guessed refunds cap at $250. In June you raised it to $500. Both are in memory now, side by side, nothing marking which is true. Guess which one it quotes.
BLOAT

Everything, every time

The common fix: inject everything it ever learned into every request. Ask about one refund and the deploy process rides along, and the 2019 policy, and last week's notes. The two lines you needed are in there, somewhere, behind everything you didn't. More memory, dumber AI.
03 · THE FIX HAS A SHAPE

A context graph is context treated as an asset.

Not a chat log, not a document pile, not a vector soup. A graph of discrete facts — each one carrying where it came from, how confident it is, and how it relates to the rest — versioned like code and owned like property. Four properties make it an asset instead of a liability:

Provenance — every fact cites itself

Who learned it, from how many sources, at what confidence. Trust becomes checkable.
"Chargebacks route through disputes" · conf 0.94 · 14 sources · claude · 7d

Curation — knowledge earns its place

New learning is proposed, not assumed. Rules auto-accept the routine; humans judge the rest. Wrong facts get retired, not buried.
merged·auto held·review retired

History — understanding has versions

What did we believe in March? What changed, and why? Versioned knowledge can be pinned, diffed, and rolled back.
v2.4.0 → main · +39 facts changed · cut Aug 13

Composition — graphs build on graphs

Your team's graph depends on the company glossary and a public industry ontology. Fluency is assembled, not retyped.
deps: finance-core @2.1 · company-glossary @1.4 · fin-ontology @3.0
f-1002conf 0.91
Disputes queue SLA is 2 business days
f-1001conf 0.85
Chargebacks route through the disputes queue, not refunds
f-1225⚠ contradiction
Refunds over $250 need director sign-off
f-1224superseded
All chargebacks are handled as standard refunds
— supports- - supersedes·· contradicts
facts relate — supporting, superseding, contradicting — and the graph keeps score
04 · THE MULTIPLIER

Owning context is the differentiator. Retrieving only what's needed is the multiplier.

A well-kept graph is only half the win. The other half is what reaches the context window: the relevant subset for this task — ranked, capped at a token budget, every line citing its source — and nothing else. Focused context makes the same model faster, cheaper, and measurably sharper. That's not an optimization. Compounded across every request your team and agents make, it's the whole game.

MEMORY AS A DUMP
everything it ever learned, every request
the two relevant lines are in there. somewhere.
MEMORY AS A GRAPH
refunds/policy · conf 0.93 · f-8f2a · merged
refunds/routing · conf 0.81 · f-c11d · your branch
terms: ARR · company-glossary @ v1.4.0
ranked · token-budgeted · provenance-tagged
what this task needs. nothing else.
ONE LOOKUP, WITH THE RECEIPT1 span matched
bump the churn rate number before the QBR
1 term · 3 definitions · 45 tokens handed to the agent semantic top-8 on this same utterance: 121 tokens

cost is part of the interface — every retrieval shows what it handed over, and what it saved

05 · WHOEVER YOU ARE

The same asset, compounding differently.

FOR YOU
Every tool you try already knows you. Your context stops being each platform's switching cost and becomes your own compounding advantage.
FOR AN ENTERPRISE
Tribal knowledge stops retiring with the people who hold it. Agents run on reviewed, auditable truth — and every new agent starts domain-fluent.
FOR A PRODUCT TEAM
What your product learns about each customer becomes a durable, governed asset — a moat made of understanding, not of hostage data.
FOR A PLATFORM
Every tenant conveys how they run their business once — structured, isolated, provable — instead of re-explaining it in every text box you ship.
AND NOW, THE PRODUCT

Semantic.ly is the context graph, built.

The federated memory layer for AI: versioned, governed knowledge graphs your agents write to and any tool reads from — commits, branches, review, and dependencies included. One MCP server connects Claude Code, Claude.ai, and anything that speaks the standard; your graph follows you everywhere.

REVIEW QUEUE
Promote chargeback flow knowledge
agents/claude → main · 43m
rule: refund-amounts → review
RETRIEVAL / 24H
3,105 context pulls · 4 connected tools · median 820 tokens
REMEMBER
Agents write what they learn to their own branch — continuously, safely, deduplicated.
GOVERN
Your policy auto-merges the routine and holds the judgment calls. Every action audited.
RECALL
Any connected tool pulls the focused, cited subset that this task needs. Nothing else.

Start compounding.

Semantic.ly is in private beta. Join the waitlist and we'll open the door in order.

early access · one email when it's your turn · nothing else