After working together for a while, an AI should repeat fewer mistakes that have already been corrected. Saving the conversation does not produce that change on its own. Feedback may sit beside the third version of a report; an important condition may be buried in a group reply. Neither will necessarily be in the input when the next task begins.
Leaving all the organization to the user adds another daily job: find useful feedback after each delivery, rewrite it as experience, and put it in the right place. Auto Dream aims to take on this work so that a role’s working history can inform its later approach.
The difficult choice is allowing experience to be saved automatically. This reduces maintenance, but an unsuitable generalization may then enter a future task. Teloa currently keeps experience with a source, a scope, and a way to withdraw it. These constraints make mistakes traceable; they do not establish that the model’s interpretation is correct.
One correction can produce two very different lessons
Consider a hypothetical research task. An employee compares two reports. The user points out that their reporting periods differ and asks for the conditions to be explained before deciding whether a comparison is useful.
That feedback could be summarized as:
- Check reporting periods before comparing, and explain how differences affect the conclusion. This is a research method.
- Never compare sources with different reporting periods. This turns the original problem into an overly absolute rule.
Both can cite the same real feedback, yet they lead to different behavior. The second may cause a later analysis to be abandoned even when it could have been completed.
The existence of evidence and the soundness of an interpretation need separate assessment. A source lets us revisit the original request; scope limits where an experience can be used. Whether it helps depends on what happens in the next task.
Why give reflection its own time
During delivery, the immediate priority is to meet the current requirements. Asking the employee to extract general methods at the same time can promote a temporary request into a lasting rule too quickly. It can also keep extending the original task.
Auto Dream makes the daily review a separate piece of work. The current mechanism reads five kinds of evidence for the role and date in question: runs, artifact versions, approval outcomes, the owner’s revisions, and group replies. The run fixes the role and date; the model uses that material to produce a digest and experience candidates.
These sources complement one another. The first draft shows the initial interpretation, revisions show what changed, approvals show which actions were actually allowed, and group replies preserve how results returned to shared work. This gives the review a more inspectable basis than the employee’s own account of what it learned.
The review run can only read the day’s evidence and submit its digest. It cannot use reflection to carry out new business actions. Separating the work leaves room for interpretation while preserving the original delivery as it happened.
A digest, saved experience, and later use are separate steps
There are three points at which the content changes:
- The daily digest records the day’s work. It can explain why a report was revised without turning every event into a future rule.
- Experience candidates extract approaches that may be useful later. A digest can currently submit up to three candidates. The service checks their source, role, and scope, then automatically saves and activates eligible entries. Suggestions to withdraw existing experience remain decisions for the owner.
- Later tasks select applicable material. Saving an experience does not send every entry into every turn. Use still depends on the role, scope, active state, and actual reading path.
Each step can fail differently. A digest that omits feedback leaves the interpretation without evidence. A candidate that generalizes too broadly may supply the wrong instruction later. A suitable experience that never reaches the turn cannot change its result. Keeping the steps separate helps locate the correction.
Auto Dream does not train model weights. It changes the material the system stores and supplies to later work, which also gives it an explicit place for maintenance.
What automatic saving costs
Requiring individual approval for every routine experience would create another review queue. A user hoping to avoid repetitive organization would instead have to approve summaries of every small correction. Teloa therefore automatically saves entries that meet its source and scope rules.
This is a trade-off between maintenance effort and the risk of faulty generalization. Current checks can establish which role an entry belongs to, which digest it came from, and where it may be used. They cannot establish whether “never compare different periods” is a sound business judgment. A semantic error may remain after source validation succeeds.
We therefore treat this mechanism as traceable automatic organization, rather than autonomous learning whose effectiveness has already been demonstrated. Its value depends on the work saved relative to the effort of finding and correcting mistakes. The number of saved entries cannot answer that question.
The choice also requires withdrawal to hold. Retrying the current automatic save does not reactivate an entry that has already been withdrawn. Processing its source again cannot use that same save process to override the owner’s decision. Otherwise, a correction would change the interface briefly while allowing the old instruction to return.
The next task tests the interpretation
When reports with different periods appear again, we want the employee to check the conditions, explain the difference, and complete an evidence-based comparison. If it refuses without judgment, the original feedback should be revisited to see whether the lesson became too broad.
The investigation can be specific. Did the relevant experience reach this turn’s input? Did it express the original requirement accurately? Did execution follow it? Did the delivery meet the current standard? Usage records answer only part of this; the result still needs its own assessment.
An unsuitable experience can be withdrawn and new content organized within an appropriate scope. The original report and feedback remain. This changes the future approach without recasting a past mistake as successful learning.
Auto Dream aims to make a correction already paid for help reduce the next round of rework. The next task determines whether that experience deserves to stay.