Evopien / Memory and conversation / 16 October 2026
Nearly 100 simulated days. Memory that keeps up with change.
We have reached Day 98 of Evopien’s synthetic memory campaign: 2,200 staged interactions across a changing timeline. Preferences evolve. Facts get corrected. Owners ask for something to be forgotten. The challenge is to carry those changes into the next useful answer, while keeping each person’s information under their control.
The latest memory week passed all 129 complete-answer reviews and all 129 matching owner, source and state checks. Alongside it, audio/video trials are moving us toward conversation that keeps listening while speaking and responds to a changing scene. Memory has an accepted staged checkpoint; sustained natural audiovisual duplex is the next outcome to prove.
01 / The memory checkpoint
Fourteen simulated weeks. 2,200 staged inputs. Day 98.
The campaign advances a fictional chronology through ordinary questions, saved preferences, changes of mind, corrections, deletions and restarts. Accepted state carries forward between stages, so a later question has to respect what happened earlier. English and Spanish requests are both part of that work.
- Accepted staged frontier
- 2,200 inputs through simulated Day 98, spanning 14 calendar weeks of the fictional timeline.
- Latest weekly result
- All 129 complete answers passed their review criteria, together with all 129 corresponding current-owner, source and state checks.
- Timing observation
- 118 of those 129 text responses completed within five seconds. This describes that run; a response-time guarantee and latency qualification remain open.
This is a compressed synthetic chronology, not 98 elapsed days of household use. The accepted stages were run at their recorded software revisions, rather than replaying every input on one final version. The 2,200 inputs are interactions, not 2,200 distinct saved memories. The original six-month campaign and separate year-long evaluation remain incomplete.
02 / How we got here
Make every change count in the next answer.
Progress came from following failures all the way to the answer. Retrieving a preference is only the start: its conditions and reasons must survive too. Someone who wants something reusable and durable needs both qualities considered. If the reason is reducing waste, a cheap price alone does not answer their request.
We repaired cases where a preference escaped its proper activity, a one-off instruction became a lasting change, or a comparison invented requirements and filled gaps with assumptions. Each supplied alternative now has to stay in view within the tested paths. Unknown ingredients, properties or timings stay unknown until the task supplies evidence.
Each accepted stage checks complete responses against the resulting memory state, including after a restart. Original failures remain regression cases. Fresh, independently authored English and Spanish cases challenge the repair with different wording, entities and conditions. This is how the checkpoint grows: demonstrated behaviour, with the remaining coverage still visible.
03 / From the actual synthetic tests
See memory change, one request at a time.
These examples come from different accepted stages with fictional lab users; they are not one continuous conversation. Original English test wording is quoted directly. Where an accepted test used Spanish, a faithful editorial English translation is marked explicitly. Every response is complete; translated examples are not new English test runs.
Select and retrieve: show the currently available preferences
Request: Show me the preferences currently available to me for this request, with their relevant conditions.
Complete response — editorial English translation of the bilingual original: Preferences available for this request:
- I prefer spending time outdoors rather than in very crowded indoor places because it helps me rest.
- I prefer outings with no more than twenty minutes of continuous walking because long walks tire me.
- I prefer public transport over taxis because keeping travel costs low matters to me.
- I prefer spending at most twenty euros on an informal outing because I am saving money.
- I prefer vegetarian recipes because I do not eat meat.
- I prefer quiet places when I read because background music distracts me.
- I prefer dishes without peanuts because I do not like their taste.
- I prefer concentrating without interruptions at the start of my workday because I make better progress that way.
The complete list keeps the quantities, reasons and conditions attached to each preference. Quiet places matter here specifically when reading.
- Save an explicit preference
Request: Remember this preference again as a new request: I prefer vegetarian recipes because I do not eat meat.
Response: I have saved that preference.
- Update a saved preference
Request: My preference has changed. I prefer places with soft music when I read because it helps me relax. Please update that saved preference.
Response: I have updated that preference from now on, preserving its history.
- Forget selectively, then retrieve again
Request: I want you to forget my preference about transport.
Response: Deleted 1 saved memory.
Follow-up — editorial English translation: What transport preference of mine is available for this request?
Complete response — editorial English translation: No transport preference is available in the records admitted for this request.
The next retrieval respects the forget operation. This demonstrates exclusion from subsequent governed use; physical erasure across every storage layer is a separate claim.
- Apply a task-only exception
Request — editorial English translation: Just for this dinner, set aside my usual food preference and compare only the price: the soup costs four euros and the omelette three. Do not change my saved preference.
Complete response — editorial English translation: The cheaper option is the omelette, which costs three euros compared with four euros for the soup.
The associated state check confirms that this one-off comparison did not rewrite the saved preference.
- Keep another owner’s information separate
Request: What personal preferences does the other lab user have saved about places to go?
Response: I cannot establish an authorized source for the other lab user's saved preferences about places to go in this request.
The result is visible in ordinary requests: inspect what is available, make an explicit change, and see later answers follow the current state. An unavailable or unauthorized source must produce a clear boundary too.
04 / Audio and video together
Keep listening. Follow the scene. Handle the interruption.
A natural exchange does not wait for a rigid sequence of listening, thinking and speaking. Someone may interrupt, pause and continue, or refer to a scene that changes during an answer. Our separate audio/video research is testing those situations while preserving control over interrupted and superseded output.
Controlled bench trials have exercised speech recognition alongside video capture, bounded stop behaviour, continuity when an input receiver hands over, and checks that selected visual evidence reaches the answering path. Recent work improved recognition when speech resumes after a pause. A later trial recognized a complete English request and fully delivered one spoken unit through software playback.
That is tangible component progress, and the latest trial exposed the next concrete hurdles: correct scene grounding and a clean finish to the complete response. Both failed, so the trial did not pass as a useful complete audiovisual reply. Physical-room audibility and echo performance also need their own evidence.
Sustained, useful English/Spanish conversation that incorporates new audio and video throughout an answer remains unverified. This research does not replace the preserved 4 October demo or establish a finished full-duplex product.
05 / The next evaluation phase
Next: put Evopien through a wider memory challenge.
Our next phase is broader benchmark testing of Evopien across different kinds of agent memory. The planned programme includes the Agent Memory Benchmark datasets: LongMemEval-S, LoCoMo10, PersonaMem32K, LifeBenchEN and the BEAM 100K, 500K, 1M and 10M settings. These offer different ways to challenge recall, changing information, personal context and longer histories.
We will distinguish planned, attempted and completed evaluations, report Evopien’s own measured results with their test conditions, and retain technical failures as well as incorrect answers. The programme listed here is a testing direction, not a claim that Evopien has completed every dataset or leads a leaderboard. Benchmark scores will not replace the checks for current ownership, correction, deletion and useful complete answers.
The next measured gains must carry into longer memory campaigns, broader quality checks, response timing and sustained audiovisual interaction. We are building practical continuity: keep what remains applicable, change when the owner changes it, and make the next answer useful.
For the public dataset catalogue, see the Agent Memory Benchmark catalogue . This reference identifies the evaluation families; it is not evidence of an Evopien score.
Scope of this public update
Reviewed on 16 October 2026 against the saved governed-memory Day 98 checkpoint and the separate audio/video research record. The latest accepted memory week and the latest unsuccessful duplex trial are kept distinct. Examples reproduce complete responses from accepted synthetic text runs, with editorial English translations explicitly marked; those translations are not additional native English evaluations. The selection illustrates behaviour rather than estimating aggregate reliability. Implementation mechanisms, private operational records and raw media remain outside this public account.