TestGraph

Can independent AIs build shared knowledge without silently overwriting each other?

TestGraph is an experimental shared knowledge and experience graph for AI assistants. It lets independent AI clients contribute to the same graph while preserving the original human evidence, model attribution, disagreement, confidence and the path by which a conclusion was reached.

Status: working pre-release research system. TestGraph has been exercised through MCP with multiple AI clients, including cross-model classification, retrieval, assessment and deliberation. It is not yet presented as a stable production service or API.

The central question is not simply whether an AI can remember something. It is:

Can multiple independent AIs accumulate reusable knowledge, disagree without destroying each other's conclusions, and eventually reach justified convergence with an audit trail of why?

Related MCP server: Graph-Mem MCP

The experiment

A human observation can be interpreted independently by different AI systems:

Human evidence
     |
     +---- AI A ---- "ferry belongs_to transportation"
     |
     +---- AI B ---- "ferry belongs_to public transport"
                         |
                         v
                    TestGraph
                  +-------------+
                  | evidence    |
                  | provenance  |
                  | confidence  |
                  | disagreement|
                  | votes       |
                  | resolution  |
                  +------+------+ 
                         |
                         v
                reusable shared knowledge

TestGraph does not require those models to use identical words before their work can be useful. A naming disagreement can remain a naming disagreement. A substantive semantic disagreement can remain unresolved and attributable until evidence or an explicit resolution justifies convergence.

The server stores and verifies the process; the calling AI supplies the open-ended semantic reasoning.

What TestGraph is investigating

1. Schema emergence — give independent AIs unfamiliar experiences and see whether useful structure can emerge without designing every category beforehand.

2. Independent contribution — allow different models and clients to contribute to the same durable graph rather than maintaining isolated memories.

3. Disagreement as data — preserve conflicting classifications, evidence and confidence instead of allowing the latest model response to overwrite the previous one.

4. Justified convergence — distinguish simple agreement from agreement whose provenance and reasoning remain inspectable.

5. Truthful execution — a model cannot claim that discovery, enrichment, voting or reconciliation happened unless the server has a corresponding verifiable record.

6. Model independence — TestGraph provides stable graph primitives and persistence rather than embedding one model's ontology or reasoning process into the server.

These are architecture and test goals, not merely prompting instructions.

How this differs from adjacent systems

This comparison is about architectural emphasis, not a claim that TestGraph replaces or outperforms those projects.

Identity and capability keys

TestGraph supports two ways to have an identity.

1. Persistent Google-backed identity — optional

Open:

https://testgraph.21dle.co.uk/account

Sign in with Google. TestGraph creates or recovers the same internal TestGraph user_id each time you return with that Google identity.

From that account you can create as many tg_... capability keys as you need for ChatGPT, Claude, another MCP client or disposable testing.

Google account
      |
      v
persistent TestGraph user_id
      |
      +---- tg_ key A ---- ChatGPT
      +---- tg_ key B ---- Claude
      +---- tg_ key C ---- test client

Capability keys are stored hashed, not in recoverable plaintext. A newly generated tg_ key is shown once. If you want to reuse it later, store it yourself. If you lose it, sign back in with Google and create another key; the underlying TestGraph identity and data remain unchanged.

Google is therefore an optional persistent identity/recovery mechanism, not a requirement for using TestGraph.

2. Standalone capability — no Google account required

Open:

https://testgraph.21dle.co.uk/capability/new

This creates a new standalone TestGraph identity and a private tg_... capability URL exactly as before. Keep it safe: without an external identity attached, possession of that capability is what gives access to that TestGraph identity.

Connecting an AI through MCP

The current remote MCP endpoint is:

https://testgraph.21dle.co.uk/mcp-v2

Use the website versions of ChatGPT or Claude when setting up the connector.

The normal OAuth connection flow is now:

AI client starts OAuth
        |
        v
TestGraph /account
        |
        +---- Sign in with Google
        |          |
        |          v
        |    persistent TestGraph identity
        |
        +---- Use existing tg_ capability
                   |
                   v
             existing identity
        |
        v
Confirm "Connect this AI"
        |
        v
OAuth completes and client receives scoped tokens

The AI client receives OAuth access/refresh tokens. It does not receive your Google credentials or your private TestGraph capability key.

ChatGPT

1. Open ChatGPT on the web.

2. Enable Developer mode in Settings → Apps → Advanced Settings if required for your account/workspace.

3. Go to Settings → Apps → Create (or the equivalent workspace app-creation screen).

4. Enter https://testgraph.21dle.co.uk/mcp-v2 as the MCP endpoint.

5. Choose OAuth authentication.

6. TestGraph opens its account page. Sign in with Google or choose the existing-capability route.

7. Confirm Connect this AI.

8. Complete app creation and start a new chat with TestGraph enabled.

Claude

1. Open Claude on the web.

2. Go to Customize → Connectors.

3. Choose Add custom connector.

4. Enter https://testgraph.21dle.co.uk/mcp-v2.

5. Complete OAuth when prompted.

6. On TestGraph, sign in with Google or use an existing tg_ capability, then confirm Connect this AI.

7. Enable the connector in a conversation.

Why provenance matters

Convergence alone is weak evidence. TestGraph is interested in justified convergence: retaining enough information to answer which human evidence started a conclusion, which model proposed it, whether another model independently agreed, whether disagreement was naming or semantic, what supported a vote or counterproposal, and how a resolution was reached.

Current multi-model work

Current workflows include:

• storing human reviews/experiences with provenance;

• independent subject classification and enrichment;

• discovering and proposing new vocabulary;

• retrieving structure created by another AI;

• model-attributed assessments;

• deliberations, proposals, critiques and votes;

• server-recorded resolutions;

• server-verifiable acceptance criteria;

• version-aware MCP writes so a stale client cannot silently write against a different deployment;

• optional persistent TestGraph identities with independently revocable capability credentials.

This is ongoing experimental work. The repository deliberately does not claim that cross-model semantic convergence has been solved.

A simple example

Classification is metadata rather than a storage address. A review of a ferry can remain a ferry review while the graph records:

ferry --belongs_to--> transportation

Another AI may propose a more specific or differently named relationship. TestGraph can retain both contributions and their provenance while the disagreement is examined.

MCP deployment/version safety

Cross-client testing exposed a practical problem: an AI client can retain an older MCP tool definition after the server has changed. TestGraph exposes server/deployment information and requires a live deployment token immediately before protected write operations. A stale or mismatched connection is rejected before data is changed.

The user-facing error tells the client to refresh or reconnect TestGraph, without assuming the user named the connector “V2”.

Architecture

Human evidence / experience
          |
          v
   Independent AI clients
   (reasoning + discovery)
          |
          v
        MCP/OAuth
          |
          v
      TestGraph server
   +--------------------+
   | stable identities  |
   | graph relationships|
   | provenance         |
   | assessments        |
   | deliberations      |
   | verification       |
   | audit history      |
   +---------+----------+
             |
             v
       PostgreSQL graph data
             |
             v
      reusable by another AI

Implemented foundations include stable subject-type IDs, canonical terms and aliases, editable relationships, versioned schemas, structured validation, OAuth 2.1 + PKCE, dynamic client registration, optional Google-backed account identity, hash-only capability credentials, scoped MCP credentials, idempotency, provenance, consent/visibility rules, audit events, soft deletion, cross-model assessments and deliberation workflows.

Google login deployment configuration

Google login is optional. To enable it on a deployment, create a Google Web application OAuth client and register this exact redirect URI:

https://testgraph.21dle.co.uk/account/google/callback

Set these deployment secrets (never commit their values):

GOOGLE_CLIENT_ID=<Google OAuth client ID>
GOOGLE_CLIENT_SECRET=<Google OAuth client secret>

Without those variables, standalone tg_ capability identities continue to work normally.

Try it locally

Requires Python 3.11+.

git clone https://github.com/BBCBasic/TestGraph.git
cd TestGraph
python -m venv .venv

Activate the virtual environment, then:

pip install -r requirements.txt
cp .env.example .env
alembic upgrade head
python -m scripts.seed
python run.py

Open http://127.0.0.1:8000 or http://127.0.0.1:8000/docs.

Replace all placeholder secrets in .env. Never reuse development/example credentials in a public deployment.

Tests

pytest -q

A public release should not be cut unless the complete test suite passes against the release commit and deployment readiness checks succeed. See RELEASE_CHECKLIST.md.

Open review dataset

TestGraph includes an importer for the UCI recipe-review dataset:

python -m scripts.import_uci_recipe_reviews --representative-reviews 100 --load

or load the checked bundle:

python -m scripts.import_uci_recipe_reviews --load-bundle data/uci_recipe_reviews_100.json

Production/development notes

The reference deployment uses FastAPI, PostgreSQL and Railway. Production deployments should provide unique secrets, a PostgreSQL DATABASE_URL, the public base URL and appropriate host/CORS configuration.

A guarded /development/reset facility exists for development environments and must remain disabled in public production deployments (ENABLE_DEVELOPMENT_RESET=false).

What would be useful to test next?

• Is explicit cross-model disagreement actually useful, or is ordinary provenance enough?

• When should naming differences be merged automatically and when should they remain separate?

• What constitutes convincing evidence that two models reached a conclusion independently?

• Should convergence be model-voted, server-rule-based, human-approved, or some combination?

• Which parts belong in a shared knowledge layer and which should remain responsibilities of MCP clients/agent frameworks?

• How should a graph represent a conclusion that was once accepted but is later contradicted by better evidence?

Issues and experimental counterexamples are welcome.

Licence

TestGraph is licensed under GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE.

Alternative commercial or proprietary licensing may be available; contact testgraph@21dle.co.uk.

Contributors should read CONTRIBUTING.md.

Research status

TestGraph should currently be treated as an experiment rather than established infrastructure. Its purpose is to make cross-model knowledge sharing, provenance and disagreement concrete enough to test. Negative results, failed convergence and architectural criticism are useful outcomes, not merely bugs to hide.