Skip to content
ISEGORIABenjamin Haire
日本語
WORLD MODEL MCP

Keep the finding.
Keep the evidence.
Keep room to revise.

I built research memory for Codex and Claude that keeps three things apart: what a source reports, what an assistant infers, and why you might change your mind.

Explore a synthetic example ↓Read the source ↗
● Local on your Mac · Human review · No hosted AI required
ONE RESEARCH CHECKPOINT◈
01 / SOURCEWhat was actually reported?Original material + a precise locator
↓
02 / CLAIMA finding, with its conditions.Scope · assessment · uncertainty
↓
03 / INFERENCEA conclusion you can inspect.Premises + explicit assumptions
✓ Human review↺ Revision history
THE IMPORTANT DISTINCTION

“Accepted” means reviewed for inclusion.
It does not mean proven true.

An accepted hypothesis remains a hypothesis. Similarity and repeated citations do not turn it into a fact, and holding that line is the reason I structured the store this way.

INTERACTIVE EXAMPLE

Watch a finding become revisable knowledge.

Follow one checkpoint across recording, human review, a fresh assistant session, and conflicting evidence.

FICTIONAL DEMONSTRATIONIn-memory browser example. No connection to a real knowledge store. Nothing is saved or uploaded.
SYNTHETIC / GENERALNot recorded

HISTORY

    The application itself uses a persistent SQLite store, independent MCP processes, and an authenticated local inspector. I wrote this example to walk through that workflow, with deliberately fictional findings.

    THE LOCAL PRODUCT

    A durable research record,
    with a useful interface.

    Browse evidence and assumptions, compare proposal changes, explore a bounded graph, and recover earlier accepted revisions.

    Read the verification report ↗
    01

    Evidence stays attached

    Source passages, locators, applicable conditions, and provenance travel with a claim.

    02

    Review comes first

    Assistants preserve pending proposals. A person reviews them before acceptance.

    03

    Disagreement stays visible

    Challenges do not silently rewrite accepted knowledge. Revision history remains inspectable.

    04

    Retrieval stays local

    English/Japanese keyword and graph search work offline. A verified multilingual model is optional.

    FROM THE RUNNING APPLICATION

    The evidence inspector.

    I captured this from the local browser workflow using labeled synthetic data.

    World Model local inspector showing a synthetic sleep finding, its population and conditions, and linked evidence.
    RUN IT ON YOUR MAC

    One local store.
    More than one assistant.

    Apple Silicon macOS. The setup script I ship installs an isolated Node.js 24 runtime and the locked dependencies. Connect Codex or Claude using generated absolute-path MCP configuration.

    Installation guide ↗Client setup ↗API and schemas ↗
    LOCAL INSTALLATION
    git clone https://github.com/theisegoria/world-model-mcp.git
cd world-model-mcp
./scripts/setup.sh
./bin/world-model-mcp open

    Optional model downloads require an explicit install action. Personal research is not imported automatically.

    What I claim for this, and what I cannot promise.

    It supplies durable, inspectable, revisable knowledge to connected assistants. It does not change model weights, guarantee truth, or guarantee that an assistant follows research instructions. I treat human review as a workflow safeguard, not as a security boundary against unrestricted programs running under the same OS account.

    I report build, protocol, browser, semantic, benchmark, and real-client evidence separately in the repository, and I name there any client integration I have not verified.