Appfy · internal engine

The agentic engine that compounds.

Alembic distills the world into verified Signals and Learnings and runs ventures across a model swarm. With the Hermes fusion it finally remembers what it learned — every run sediments validated lessons that make the next run smarter.

How the loop closes
Engine for FounderOS + every Appfy venture · not a sellable SKU (ADR-0001)
7
shipped subsystems
565
cases green
$0 → frontier
4-tier funnel T0→T3
fail-closed
gate at every paid tier
The problem the fusion solves

An engine that was powerful, but amnesiac.

Alembic had two of the three pieces a self-improving system needs — it could ingest the world and distill it into value. It was missing the third: nothing read its Learnings back to change how the next run reasoned. A lesson learned in run N was invisible to run N+1.

before · the dead end

The Learnings store filled up, but nothing fed it back into the engine. Each run started cold.

SOURCE → distill (T0–T3)
  → Learnings + Signals
  → [ no feedback into the engine ]
after · the closed loop

The Hermes loop slots into machinery Alembic already had — a Validator-gated return path that sediments only validated wins.

SOURCE → distill → Learnings
run → review fork ┘→ memory + skills
  → next run is smarter
The differentiator

Propose → gate → sediment. Learn only from validated wins.

Hermes auto-wrote to its memory. Alembic changed the one thing that matters for an engine: the reviewer proposes, the existing Validator disposes. Nothing sediments without clearing the quality floor (ADR-0018, default scoreThresholdGate(0.7)).

runfinishes a unit ReviewProposerproposes writes Validator Gateapprove · reject memory + skillssediments curatoractive→stale→archived rejected · never written frozen snapshot → next run reads it (cache-safe)
propose → Validator gate → sediment (frozen snapshot) → next run · curator keeps the library from rotting (never deletes)

Seven subsystems, one package.

The fusion lives in @alembic/hermes, depending only on contracts + etl — the bottom of the dependency graph. Each obeys the engine's invariants: typed Result at every boundary, FsPort for IO, no throws in library code.

memory/clone

Frozen-snapshot curated memory, injected at session start so mid-session writes never invalidate the prompt cache.

learning/adapt

The closed loop, reshaped to propose → gate → dispose. The differentiator.

clarify/clone

A blocking human-in-the-loop gateway — the T4 park can now pose a concrete question and wait.

web/clone

web_search / web_extract over an injected backend, with optional LLM content compression.

skills/clone

An FsPort skill store — tiered list / view + CRUD, with path safety and a dep-free frontmatter parser.

curator/clone

A deterministic active→stale→archived lifecycle over agent-created skills. Never deletes.

media/adapt

transcribe / analyzeImage over injected cloud backends, so audio + image sources enter the funnel.

Every change keeps the baseline green.

The engine fails closed. The whole suite — typecheck, build, test, and the impeccable design gate — is the baseline every change must keep at exit 0.

alembic — baseline
$ pnpm -r typecheck && pnpm -r build && pnpm -w test
 @alembic/contracts  typecheck ✓  build ✓
 @alembic/hermes     typecheck ✓  build ✓   7 subsystems
 test files 63 passed   cases 565 passed   (0 failed)
$ npx impeccable detect docs/
✓ 0 findings — design gate clean

The engine has consumers, not customers.

Per ADR-0001 the engine is internal infrastructure. The capability set flows to the products and ventures that sit on it — and, uniquely, to the engine itself.

first sellable product
FounderOS
Its agentic features get sharper over time without FounderOS shipping new model code — the improvement happens in the shared engine.
the portfolio
Appfy ventures
Each venture is built by the Venture Factory on the engine, inheriting durable memory + a Validator-gated learning loop.
self-consumer
Alembic itself
The Learning value-chain's whole purpose is engine self-improvement — the loop lets the engine act on the Learnings it distills.

An asset that appreciates, not a static tool.

The engine stops being something the ventures call and becomes something that gets better the more it is used.

Re-read the loop