Request 1
A small evaluation set
Questions, the sources that answer them and decision revisions a system should reproduce. Public, non-sensitive examples.
Research · Scientific AI and memory
AI that remembers why.
Intelligence compounds when nothing learned is lost. We are building DOS, a memory for human and artificial minds working together: every source kept, every decision explained, every lesson available to the next experiment. Answers you can trace, decisions you can revisit, and work you can verify.
Capability-specific development · updated September 28, 2026 In daily use by the founder with AI collaborators; each capability listed with its real scope and date.
The idea
Memory that makes intelligence compound
The public question. Can a research system answer what changed, show where that answer came from and preserve what a human decided to do about it?
DOS is the memory MNEOS runs on: a system for keeping sources, decisions, context and work across every project, developed under NuvoNexus, LLC and proven first in the founder's own daily work with AI collaborators. The goal is an institution whose intelligence grows with every project, human and artificial minds learning together, with people always able to check the record.
Illustrative example
One record: source, conclusion, decision, new evidence, revision
- Source
- A supplier datasheet for a candidate bench sensor, filed with its retrieval date.
- Tentative conclusion
- The sensor's stated response time appears adequate for the planned measurement window.
- Decision
- Include the sensor in the first bench configuration; record the assumption that the datasheet figure holds at bench conditions.
- New evidence
- A published characterization note reports slower response at the relevant operating temperature.
- Revision
- The assumption is marked superseded, the decision is reopened, and a bench timing test is added to the characterization plan. Both the original and revised records remain readable, with dates and the person responsible.
Illustrative record built from the kind of decision the bench work involves. No private correspondence, relationship data or security configuration is shown; this is not a live query.
DOS capability table
Where each capability stands today
Each capability with its real scope, stage, evidence and date. This is the honest map of a system being built in the open.
| Capability | Scope | Stage | Evidence record | Last verified | Limitation |
|---|---|---|---|---|---|
| Source ingestion | Documents, notes and session records from the founder's projects | Development | Working in the founder's daily workflow for text sources; evidence review of a public example pending | September 28, 2026 | Text and Markdown sources only; no automatic capture of private messaging |
| Retrieval with citations | Answer a question and point to the record it came from | Development | A retrieval service exists; keyword ranking is the validated lane, hybrid ranking is not yet validated | September 28, 2026 | Readiness checks that exercise retrieval are still being added; a green health check does not prove retrieval works |
| Decision and version history | Keep why a decision was made and how it was later revised | Development | Decision records are kept as versioned text in private Git-backed stores | September 28, 2026 | End-to-end review of a decision-to-revision chain is pending; see the illustrative example below |
| Communications and attachment ingestion | Bring email and attachments into the record with consent and scope | Development | Email triage runs through authorized connectors in the founder's workflow | September 28, 2026 | Message-capture bridges for other channels are not operational; no LinkedIn or iMessage capture is claimed |
| Relationship mapping | Map people, organizations and projects to the evidence that connects them | Development; evidence review pending | A graph component exists in development | September 28, 2026 | Graph ranking is not validated; no public relationship graph is exposed |
| Task execution verification | Distinguish a completed task from a plausible report of completion | Design | Scheduled checks currently verify transport and process health as separate signals | September 28, 2026 | Outcome verification is designed, not demonstrated |
| Permission boundaries | Who may know, decide, disclose, approve or act | Design | Roles and disclosure classes are documented in the institutional architecture | September 28, 2026 | Enforcement inside the software is pending |
| Backup and restore | Recover the record after loss | Development | Automated snapshots to private repositories run on a schedule | September 28, 2026 | Restore drills have not consistently passed as of September 2026; snapshots are not treated as proof of recoverability |
Evaluation and integration
How we would know it works
Evaluation questions we are working on
- Given a question, does the system return the record the answer came from, not a plausible passage that resembles it?
- When a conclusion is revised, do both the old and new conclusions remain legible, with dates and the responsible person?
- Can the system tell a completed task from a report that a task was completed?
- Does a green health check correspond to working retrieval? (It did not always; readiness checks that exercise retrieval are being added.)
- Which sources should never enter the record without explicit consent and scope?
Current integration work
- Retrieval readiness probes that validate the index before serving answers.
- Rebalancing hybrid ranking so graph facts do not overwhelm ordinary relevance; keyword ranking remains the validated lane meanwhile.
- Restore drills for the backup path, since scheduled snapshots alone do not prove recoverability.
- Structured intake for operational tasks so a request, its owner and its evidence of completion are recorded together.
Join the build
Three ways to shape DOS now
Request 2
A revision representation
A proposal for how a superseded conclusion and its replacement should both stay readable and retrievable.
Request 3
Retrieval scoring that rewards the source
A scoring method that penalizes plausible-but-wrong passages and rewards returning the actual record.
Open collaboration. Scope, time, access and compensation are agreed before work begins.