An AI second brain can remember everything and still be wrong

Old facts return, notes disagree, and agents write faster than people can review. Search finds what looks relevant without knowing whether it is current or supported.

The structure shown below addresses this through eight connected layers:

CORE: current project facts
DOCS: models, assumptions, and scope
SRC: code that runs
DECISIONS: what changed and why
EVIDENCE: results and provenance
NOTEBOOKS: exploration kept separate
TESTS: automated checks
SCRIPTS: repeatable experiments

A dark network graph grouping project records into eight labeled layers: Core, Docs, Source, Decisions, Evidence, Notebooks, Tests, and Scripts — A derived map of connected project records. It helps with navigation but does not decide which record is true.

These methods are not new. Their value comes from how they connect.

A key parameter can be traced from its source and decision record to the value in the code, the check that verifies it, and the experiment behind the result. The graph shows this path without deciding what is true.

Agents propose. Evidence, checks, and people decide. A failed check blocks acceptance until the mismatch is resolved. Hallucinations still happen, but the model has less to guess and errors are easier to trace.

If you are building a second brain:

Give each changing fact one owner.
Keep replaced decisions.
Link claims to code, checks, and evidence.
Use search to find records, not to decide what is true.
Add the vector database last.

Which rule is missing from your system?