Memory, Evidence, and Bounded Inference
Status · Active research
Can connected personal evidence support responsible inference while preserving privacy, uncertainty, provenance, and human authority?
Why the question matters
Long-term collaboration requires memory, but remembered material can be copied, summarized, misattributed, repeated, or taken outside its context. The archive must preserve source lineage before it supports interpretation.
How the investigation began
The work began by preserving conversations and authored material as primary sources, then separating transcripts, evidence records, derived guidance, and human confirmation.
Human–LLM collaboration method
Canonical source material is preserved first. Deterministic extraction, provenance, contradiction tracking, temporal weighting, privacy boundaries, and human review constrain any later inference.
Projects and experiments
- Alexandria Chat Archive
Verified accomplishments
- A deterministic archive and canonical transcript workflow have been implemented.
- Search and provenance preserve routes back to source conversations.
- Safeguards distinguish primary evidence from copied reports, summaries, and derivative claims.
- A bounded psychological-layer harness and comparison console isolate authorized context for evaluation.
What has not been demonstrated
Governance has been demonstrated. Psychological usefulness has not. Archive integrity does not validate an inference, and no personal psychological details are part of this public record.
Failures, ambiguity, and competing explanations
Source completeness, duplicate material, changing beliefs, context boundaries, and derivative repetition can all distort an inference. Human confirmation cannot repair missing primary evidence.
Selected articles
- What Most People Actually Do With AI — And Why It Matters
- Do We Need AI-Integrated OS for a Smarter Future?
Selected podcasts
These episodes show continuity in the questions; they do not validate the software experiments.
Current position
Memory should improve retrieval and continuity without becoming an authority over the person whose record it describes.
Open questions
- Can bounded context improve a task without unwanted personalization?
- How should old evidence be revised without erasing its historical role?
Next bounded test
Run the predeclared five-task usefulness comparison with only authorized, context-matching records and score relevance, leakage, overreach, and material benefit.