Conversation history
What was said. Useful for continuity, but not automatically factual, current, or authorized.
OVERLIFT · EVIDENCE, MEMORY + TEMPORAL TRUTH
Remembering is not proving.
A trustworthy agentic system must preserve what happened, what a source reported, what was actually knowable, what a model inferred, what a person accepted, and what became official—without collapsing them into one convenient “memory.”
New evidence may change today’s conclusion. It must not silently rewrite yesterday’s decision record.
MEMORY HAS TYPES
Eight records that must never become one undifferentiated pile.
The same sentence can be useful as conversation context, unsafe as evidence, provisional as interpretation, and invalid as official state. OverLift classifies every retained item before allowing it to influence a consequential result.
Working state
Plans, intermediate calculations, tool results, and incomplete candidate paths.
Preferences and consent
Explicit and inferred preferences carry scope, confidence, provenance, expiry, and revocation.
Source evidence
What a named source actually reported, including version, rights, retrieval time, and correction history.
Admitted evidence
The subset that passed time, scope, quality, rights, security, and policy gates for this exact decision.
Learned interpretation
Summaries, rankings, predictions, explanations, and generated drafts remain distinct from source fact.
Official state
The artifact or state transition accepted by exact software, policy, and accountable people.
Receipt and replay state
Evidence identity, exclusions, versions, decisions, hashes, and recovery needed to reproduce the result.
FOUR CLOCKS
“When” is not one timestamp.
Temporal truth becomes reliable only when the system can distinguish the time a fact applied from the time it was observed, admitted, and used in an authoritative decision.
When the event, record, preference, or condition applied in the world.
When a source or system first encountered that information.
When the evidence passed the gates required for this decision.
The exact cutoff used to produce, accept, and seal the result.
DETERMINISTIC TEMPORAL-TRUTH LAB
What did the system know—and what was it allowed to conclude—at that exact moment?
Conversation, preference, evidence, interpretation, and official state are not the same kind of memory.
Track when something applied, when it was seen, when it was admitted, and when the decision was committed.
Newer, clearer, or more confident output remains a candidate until the governing contract accepts it.
Reproduce the evidence, exclusions, versions, policy, calculation, and receipt—not merely a similar answer.
Loading deterministic scenario…
System result
Authority boundary
Typed receipt
Memory and authority rule
- Official state changed
PRODUCT PROOF
One temporal contract, six very different domains.
The information changes. The authority rule does not: classify it, place it in time, preserve its source and version, admit it deliberately, and reproduce the accepted result later.
What was knowable then
Reconstruct historical evidence, preserve late arrivals, expose disagreements, and replay the exact research conclusion.
See Saros proof →Calculated change versus new language
Keep deterministic astronomical state separate from symbolic interpretation and longitudinal reading history.
See Veil proof →Preference is not permission
Track freshness, confidence, scope, consent, revocation, and current safety before any social action.
See Veluris proof →Candidate draft versus accepted meaning
Preserve author identity, exact document versions, protected claims, reversible diffs, and deliberate commit.
See Vellucent proof →Generated explanation versus reviewed concept
Use models to improve learning without silently replacing source-reviewed canonical knowledge.
See Glossary proof →Decision-time operational truth
Preserve source state, approvals, action identity, postconditions, and receipts for operational decisions.
See Logistics proof →