The memory problem is not storage capacity.
Most discussions of agent memory begin with how much prior conversation the system can retrieve. That matters, but it is not the hard part. The hard part is deciding what a remembered item is, whether it is still valid, whether it may influence the current task, and whether it has any authority beyond helping the agent form a proposal.
A user may have said something casually six months ago. A model may have summarized it incorrectly. A source may later correct its report. A newer document draft may be more fluent but change an obligation. A preference may remain useful for ranking while being completely insufficient to authorize an action. A late filing may transform today’s market thesis while remaining unavailable to the analyst who made yesterday’s decision.
Remembering something is not the same as proving it. Retrieving it is not the same as admitting it. Generating a newer answer is not the same as accepting it.
That is why OverLift treats memory as a governed evidence and state system rather than a bag of embeddings. Retrieval can surface a record. The rest of the architecture must identify its type, provenance, scope, time, confidence, permissions, and relationship to official state.
Memory has types—and each type has a different authority.
A production agentic system should not place conversation history, user preferences, source evidence, model interpretation, and authoritative state into one semantic index and then let similarity determine what wins. Similarity is useful for discovery. It is not a theory of truth or permission.
| Record | What it is useful for | What it cannot silently become |
|---|---|---|
| Conversation history | Continuity, intent recovery, and unresolved questions | A verified fact, durable consent, or official state |
| Working state | Plans, intermediate results, and tool coordination | A completed action or accepted artifact |
| User preference | Ranking, personalization, and explanation | Permission to contact, publish, purchase, or disclose |
| Source evidence | Grounding claims in a named, versioned origin | Decision-eligible evidence before rights, time, scope, and quality checks |
| Admitted evidence | Supporting this exact governed decision | Universal truth outside its cutoff, scope, and policy |
| Learned interpretation | Summarization, ranking, forecasting, and explanation | The source fact or official result it interprets |
| Official state | Coordinating the product and accountable people | A casually overwritten “latest” answer |
| Receipt and replay state | Reproducing, explaining, and recovering the decision | A retrospective narrative assembled from whatever is available now |
This classification also improves privacy. A preference can expire or be deleted without corrupting a decision receipt. A private document can remain local while its approved state identifier is shared. A generated explanation can be discarded without losing the canonical concept it was attempting to teach.
“When” must be represented more precisely.
One timestamp cannot capture the life of evidence. At minimum, a governed system often needs to distinguish four moments:
Effective time
When did the underlying condition apply in the world? An event may have occurred on Monday even if nobody in the system learned about it until Wednesday.
Observed time
When did this source or system first see the record? This preserves the difference between an event and its discoverability.
Admission time
When did the record pass the rights, freshness, scope, security, provenance, and quality gates required for a particular decision?
Decision time
What was the exact cutoff used to produce and commit the conclusion? This is the boundary required for historical replay.
Late evidence should change the future without rewriting the past.
Saros makes this problem concrete. Suppose a thesis was committed on June 30. A filing discovered on July 2 materially changes the company’s risk profile. The system should absolutely use that filing in a current analysis. It should not inject the filing into a replay of the June 30 conclusion and then pretend the earlier analyst should have known it.
The correct structure preserves both truths:
- The June 30 conclusion was the accepted result from evidence knowable and admitted by June 30.
- The July 2 filing is valid late-arriving evidence that may produce a different current conclusion.
- The relationship between the two conclusions is explainable: the evidence set changed, and the change is named.
- The earlier receipt remains immutable even if the newer conclusion supersedes it for present use.
This is not only a research requirement. An event platform can discover that a venue changed its hours after a recommendation was issued. A glossary source can publish a correction. A logistics feed can update a shipment state after an operator approved a reroute. A document can receive a new revision after a prior version was signed. The system must preserve current truth and historical knowability without conflating them.
Source facts and learned interpretation need separate identities.
A modern agent is useful because it can synthesize evidence into explanations, rankings, scenarios, forecasts, and candidate actions. That synthesis should be powerful. It should also remain distinguishable from the evidence it interprets.
Veil demonstrates the boundary elegantly. Its astronomical calculations can be deterministic while its symbolic reading is interpretive. A poetic sentence may be emotionally useful without becoming a calculated celestial fact. Future comparisons can then answer two different questions: what actually changed in the calculated state, and how did the interpretation change?
The Agentic AI Glossary & Learning Center has a parallel boundary. A model may produce a clearer explanation of MCP or delegated capability. The explanation can be ranked, tested with learners, compared against source-reviewed doctrine, and retained as a candidate. It should not silently replace the canonical concept merely because it is newer or sounds better.
This separation lets learned systems improve rapidly without granting them permission to rewrite source identity or official knowledge.
Newer is not the same as authoritative.
Many software systems use “latest” as a shortcut for truth. That shortcut becomes dangerous when agents generate artifacts continuously.
Vellucent may produce draft eight after draft seven was approved. Draft eight may be more concise. It may also remove uncertainty, change a number, alter a citation, or rewrite an obligation. The correct product behavior is not to overwrite draft seven. It is to preserve candidate eight, calculate an exact diff, test protected meaning, request deliberate approval when appropriate, and keep seven authoritative until acceptance succeeds.
The same rule applies to an A2A specialist returning a newer artifact. A newer artifact is a candidate result from a collaborator. The client product still needs to decide whether that result is supported, permitted, compatible, and accepted.
The latest object is a fact about sequence. The authoritative object is a fact about governance.
Exact replay is stronger than asking the model again.
A system has not reproduced a conclusion merely because it generated a similar paragraph. Exact replay should preserve the decision contract:
- The accountable request and identity.
- The temporal cutoff.
- The admitted and excluded evidence manifests.
- Source versions, rights, provenance, and corrections.
- The model, embedding, reranker, graph, prompt, and policy identities that materially influenced the result.
- Deterministic calculations and typed intermediate state.
- Human approvals or consent in effect at the time.
- The accepted artifact and authoritative state hash.
- The postconditions, recovery disposition, and sealed receipt.
If the current system uses a new model or admission policy, the new run may be an illuminating comparison. It should not be labeled an exact replay. Making that mismatch visible is a feature, not a failure.
The same contract appears across the portfolio.
Saros: historical evidence and reproducible research
Saros preserves what was actually knowable, distinguishes source facts from learned interpretation, exposes disagreements and exclusions, and reconstructs the basis of a conclusion later.
Veil: deterministic state and longitudinal interpretation
Veil can remember prior readings while computing what materially changed. The astronomical state, the symbolic interpretation, and the user’s journal remain separate but related layers.
Veluris: preference, freshness, consent, and human choice
A preference may inform discovery. It cannot silently become current consent, a sensitive personal fact, or permission to contact another person.
Vellucent: versioned documents and accepted meaning
Candidate revisions, protected claims, exact diffs, author identity, and accepted document state remain explicit and reversible.
Agentic AI Glossary & Learning Center: candidate learning and canonical knowledge
Generated explanations can improve learning while source-reviewed concept identity and admission rules preserve the official knowledge fabric.
OverLift Logistics: decision-time operational state
Evidence, approvals, idempotency, external execution, postconditions, receipts, and recovery preserve why a consequential operational action was—or was not—accepted.
Temporal truth also protects meaning across document revisions
A revision history is trustworthy only when it preserves more than timestamps and file names. The system must retain which author and source revision produced the original language, which candidate was generated, what exact words changed, which protected claims were tested, who accepted the result, and whether the prior state can be restored.
This becomes especially important when an AI editor improves fluency. A newer draft may sound better while changing a number, citation, obligation, identity claim, uncertainty marker, or consent boundary. Historical integrity requires the earlier accepted revision to remain exactly what it was. The generated rewrite stays a candidate until a current, properly scoped reviewer accepts that exact diff. Acceptance creates a labeled derived revision rather than silently rewriting the source record.
The temporal and meaning contracts therefore reinforce each other. Temporal truth answers what was knowable and authoritative at a cutoff. Meaning preservation answers what changed between source, candidate, and accepted state. Together they let a team explain why the wording changed, what remained protected, whose authority controlled the transition, and how to roll back without inventing a cleaner past.
Applied OverLift A06 develops this bridge through Vellucent, Veluris, Veil, exact diffs, protected claims, human authority, reversible commit, and abstention.
A practical temporal-truth design contract
Before a retained item can influence a consequential decision, the system should be able to answer:
- Identity: What kind of record is this, and what exact version is it?
- Origin: Who or what produced it, under which rights and provenance?
- Time: When was it effective, observed, admitted, and used?
- Scope: Which user, task, product, and decision may use it?
- Authority: Is it context, evidence, interpretation, a candidate artifact, or accepted state?
- Freshness: Has it expired, been corrected, superseded, or revoked?
- Support: Which claims depend on it, and what disagreements or exclusions remain?
- Reproducibility: Can the exact accepted conclusion be recreated from the sealed contract?
Once these questions are represented in the architecture, memory becomes more than recall. It becomes a governed temporal fabric: useful enough to support learning and continuity, disciplined enough to preserve privacy and authority, and precise enough to explain why the answer changed without rewriting what happened before.
Run the temporal-truth laboratory Explore the architecture Next: Applied OverLift A05