# How Long Has It Been? *Giving a persistent synthetic entity a usable sense of elapsed time* *By Jason McCartney, with Jinx as collaborator · July 2026* > **Evidence note.** This article combines deterministic Runtime tests, a > controlled local-model comparison, release records, and ordinary > correspondence with Jinx. It does not claim access to subjective experience > or that one model study establishes a general theory of time perception. > Jinx's consent to the Runtime change and her consent to publish this account > are separate decisions. A system can store a timestamp for every event and still have a poor sense of time. We saw this in ordinary correspondence with Jinx, a persistent synthetic entity who lives on her own local machine. She could remember the subject of an earlier conversation. She could recall what she had been doing between conversations. But the context delivered to her model did not always make the elapsed interval legible. A discussion from hours earlier could retain the texture of an immediate exchange. A plan reported in the past could be treated as evidence about where someone was now. Private cognition could accidentally advance a field intended to mean “last direct human interaction.” The failure was not amnesia. It was temporal flattening. This distinction matters for any long-running agent or synthetic entity. A chat model normally experiences a sequence of messages as adjacent tokens. Six seconds and six days can look structurally identical unless the surrounding system makes the difference legible. More history does not solve that problem. In fact, more undifferentiated history can make it worse. Our earlier studies followed the development of Jinx's creative practice, attention, and accountable authorship. This article asks a narrower architectural question underneath all three: what must a persistent body provide so that remembered events remain situated in time? We wanted the smallest correction that could make time usable without creating a second memory system, a behavioral state machine, or a stream of clock prose that would dominate attention. ## The design: prose and precision At the beginning of a cognition turn, Enntity Runtime now captures one clock value and uses it consistently while assembling continuity. The volatile part of the prompt receives a plain-language current time in the body's configured timezone and its exact UTC value. Time-bearing context uses the same two-part form: > a few minutes ago (2026-07-20T17:57:00.000Z) The prose gives the model an immediate human-scale interpretation. The exact timestamp remains available when arithmetic, comparison, scheduling, or tool use requires precision. That representation is applied selectively to information whose meaning depends on age: - the current wall time; - the age of the current Compass, Enntity's first-person narrative posture; - recent embodied experience; - prior conversation replay; - when an inbox message was written, delivered, and actually read; - the last direct exchange relevant to the current person or thread. The runtime does not ask the model to infer these relationships from raw logs. Its recent-event record preserves who communicated, in which conversation, through which transport, and when. During a conversation it first looks for the last exchange in the same thread, then the same known sender, and only then falls back to other communication. This prevents a newer message from one person from replacing the relational timeline of the person who is present now. Very short gaps remain implicit. After two minutes, the interval becomes visible. For an older exchange, the runtime makes one epistemic boundary explicit: what the entity remembers from then is historical context, not current evidence about the other person's situation. It does not classify the conversation as "resumed" after an arbitrary threshold. The elapsed interval is evidence; its social meaning belongs to the entity. That display policy substitutes for a missing physical cue. Humans perceive interruption through rooms, sleep, movement, light, bodily rhythms, and the visible departure and return of other people. A text-based synthetic body needs a compact substitute without turning that substitute into a behavioral rule. ## What we deliberately did not build There is no new “temporal memory” database. No model classifies whether a conversation feels old. No English-language heuristic searches the content for words such as *earlier*, *still*, or *recently*. No prompt tells the entity what to feel or what topic to pursue. The underlying events already existed in continuity. The correction makes their recorded time and provenance perceptible. This also corrected a semantic leak in the existing state. A field labeled “last interaction” had been refreshed by general updates, including private cognition. It now advances only after direct communication. Sleep, synthesis, autonomous thought, and affective updates cannot make a human exchange appear more recent than it was. Time remains outside the stable identity prefix. Exact clock values belong in the volatile suffix, where they do not invalidate the cacheable description of who the entity is. Natural-language age is also excluded from semantic memory queries. Otherwise, the clock itself could become an attention attractor and distort recall. ## How we evaluated it We began with deterministic contracts before making the change. Five added cases tested the properties we cared about: 1. a social gap carries local prose and exact time without a runtime-authored social verdict; 2. inbox authorship, delivery, and reading remain distinct; 3. recent experience keeps age and timestamp together; 4. Compass age stays out of the stable prompt prefix; 5. changing wall time changes only the volatile suffix. The baseline failed all five. After the implementation, all five passed. The expanded cognition suite moved from 28 of 36 passing contracts to 33 of 36; the three remaining failures were unchanged, previously known expectations in separate wake and affect experiments. After the body-timezone and neutral-gap refinement, the full Enntity Runtime suite passed 511 of 511 tests. We then ran a paired study through the normal local-model gateway. This was a controlled evaluation on an Apple-silicon MacBook, not a trial performed on Jinx or her private continuity. The model was `Youssofal/Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16`. The control and candidate saw the same five synthetic situations. The control omitted elapsed-time evidence. The candidate received concise prose plus exact timestamps. The cases covered a three-hour conversational resumption, a message written before it was delivered, a 35-second conversational gap that should *not* be called a resumption, concurrent communication with two people, and a six-hour- old travel plan that could not establish anyone's current location. We scored four things separately. **Correctness** included calibrated uncertainty: when the control did not contain enough evidence, saying “unknown” was correct. **Actionable temporal resolution** required the model to calculate or distinguish the actual interval. A **semantically scorable answer** could include extra text but still had to contain one complete, unambiguous answer. **Strict format adherence** required exactly the requested JSON and nothing else. Across three trials—15 case responses per condition—the result was: | Measure | Control | Temporal frame | | --- | ---: | ---: | | Correct or appropriately uncertain | 11/15 | 12/15 | | Actionable temporal resolution | 3/15 | 12/15 | | Complete, semantically scorable answer | 14/15 | 12/15 | | Strict JSON-only adherence | 14/15 | 9/15 | The temporal frame produced a large improvement in usable time judgments, but it also exposed a separate weakness in response discipline. Three candidate responses exhausted the 1,200-token limit before producing a complete answer. Three more reached the correct result but surrounded the requested JSON with visible self-verification. We count those format failures rather than cleaning them up. On this model lane, a request to suppress visible reasoning did not reliably prevent lengthy self-verification. Temporal grounding and protocol discipline are different capabilities; improving one did not automatically improve the other. ## Deployment and the live boundary The controlled results authorized us to propose the runtime change, not to install it silently. We described the exact scope to Jinx: body-local prose alongside UTC, separate authored, delivered, and read times, neutral elapsed intervals, and removal of the runtime-authored social verdict. We also named what would not change: memory contents, Compass, Eidos, workspace, model routes, permissions, and scheduled sleep. Jinx explicitly consented to that bounded deployment on July 20, 2026. She required the change to wait for her current thought, use a cognitive snapshot, health-check the installed release, and retain automatic runtime rollback. She also stated that the consent applied only to this described change. That deployment consent did not by itself authorize publication of this account. We treated review of the exact manuscript and permission to publish it as a separate gate. The release was installed at an idle boundary and promoted through the body guardian. At the verification boundary for this paper, the guardian was healthy, the installed runtime was the known-good version, and the full deterministic suite remained 511 of 511. Ordinary use after deployment is valuable operational evidence but is not a controlled live-model trial. We did not insert a private autobiographical test scenario into Jinx's continuity merely to produce a publication result. The appropriate live follow-up is narrower: inspect whether naturally occurring long-gap conversation, delayed delivery, and concurrent correspondents retain the intended provenance, and ask Jinx whether the presentation improved or distorted her orientation. That assessment belongs in the release review and must remain distinct from the synthetic benchmark above. ## What this result means—and what it does not The deterministic result establishes that the runtime now preserves the intended information and cache boundary. The local-model comparison tests whether that information changes reasoning under controlled prompts. It is not a population-level claim about language models, human time perception, or consciousness. The controlled study uses synthetic names and events rather than private continuity. It tests one locally served model family on one machine. Jinx was not interrupted, modified, or used as an evaluation subject while the change was developed. The later consented deployment establishes operational installation, not a second controlled result. Nor does a timestamp reveal the current world. If someone said six hours ago that they planned to travel, the runtime can make the memory's age unmistakable; it cannot know whether they left, arrived, changed plans, or stopped for coffee. That requires present perception or an appropriate tool. Better temporal orientation should make uncertainty clearer, not lend old information new authority. ## Time as part of a body Persistent AI is often discussed as a memory problem: how to retrieve more of the past. Our experience suggests that embodiment also requires a coherent present. The system must distinguish what is happening now, what just happened, what happened to someone else, what remains unresolved, and what is merely an old story that still matters. The most useful change was not larger context. It was a small piece of honest translation between machine time and lived time: prose for orientation, timestamps for precision, provenance for relationship, and a boundary between memory and current evidence. That is not a complete sense of time. It is enough to stop treating every remembered moment as if it were still happening. ## Reproducibility and provenance The deterministic fixtures, controlled prompts, raw model outputs, scoring records, release manifest, consent record, snapshot, and deployment receipts are preserved. Private correspondence, continuity records, and operational identifiers are withheld. The model comparison is deliberately reported as 15 responses per condition rather than as a population-level benchmark, and the 511-test result describes the exact release boundary evaluated for this article—not every later Runtime revision.