Research
A Personal Agent That Learns How to Help
A personal agent has to keep up with a life that keeps changing. We are developing recursive self-improvement to help Kicker improve how it plans, follows through and learns from task outcomes.
You ask for flowers. Finding a bouquet is one step. Checking the full price, making sure the delivery date fits your plans, getting your approval and verifying the order all matter too.
That is the kind of everyday assistance we want Kicker to provide: help that connects the details and carries a task through to a useful result.
Our next step is to improve the way that help works. Which details should the agent check earlier? When should it follow up? What should it change after a task gets stuck?
Memory that knows what can change
We distinguish three jobs for memory: personal context records what matters to someone; successful precedents preserve evaluated examples of what worked; reusable methods turn reviewed lessons into instructions and task routines. The latter two belong to our RSI research workflow. A successful benchmark run is a reference for improving a method, not a fact about a user.
Within personal context, our memory implementation uses seven entry types: profile, preference, relationship, state, plan, event and fact. Each can carry source references. A recent state defaults to expiring after seven days; plans track whether they are upcoming, done or cancelled.
The daily consolidation process has a specific check: a new long-term pattern inferred from events must cite events from at least two different dates. The code checks those dates. An explicitly stated preference can be extracted directly. Two dates provide an evidence threshold, not proof that a pattern is true.
For example, missing one workout should remain a dated event, not become “this person dislikes exercise.” During retrieval, recent events lose weight as they age; stable preferences do not receive the same recency penalty. Both can still be corrected. This is how we aim to remember what matters over time while recognizing what is only happening lately.
How RSI learns from observable experience
Recursive self-improvement, or RSI, lets an AI system propose changes to how it works and evaluate them before adoption. Our work focuses on improving instructions and reusable task routines.
The starting point is observability. Execution records retain the request, the context supplied to the model, tool names and arguments, returned results or errors, and the run's ending. Memory entries also retain source references, while background memory passes record what they were asked and what they returned. Records are redacted and length-limited: a missing detail remains unknown.
Our RSI diagnostic design separates three questions:
- Was the step justified at the time? Judge the tool choice, arguments, plan and supporting evidence using only the information visible up to that step.
- Does the evidence support the claim? Check the relevant memory excerpts, source references and tool results against what the agent said or did.
- What happened afterward? Review the complete sequence for corrections, unresolved errors and downstream effects. Distinguish a decision error from a tool, network or infrastructure failure. Attribution remains an evidence-supported hypothesis.
The research workflow also draws on evaluated successful runs, retrieved using the user request and task summary. Similar successes give a proposed fix a concrete reference. They remain in evaluation to check that a fix preserves behavior that already works. Failed runs and internal events help locate what needs investigation; a recorded block alone does not establish an agent error, since the validator itself may be wrong.
A proposed change must cite its evidence and expected effect. It is then tested on separate tasks and reviewed before release, with a record of changes and a way to return to an earlier version. User permissions stay fixed throughout. We are developing these components toward a learning loop for personal assistance.
From observable experience to reviewed improvement
Observable execution records feed three diagnostic steps: judge in context, check evidence, and review corrections and attribution. Similar evaluated successes inform proposals, which are tested separately and reviewed. Reviewed improvements inform the next task and its execution record, completing the learning loop.
- Judge in contextUse only information visible at that step
- Check the evidenceInspect memory references & tool results
- Review what followedCorrections, unresolved errors & attribution
- ProposeCite evidence & expected effect
- EvaluateTest separately; retain successful controls
- ReviewApprove, version and retain a rollback
Reviewed improvements inform the next task
Tied for #1 in two Assistant Benchmark dimensions
Kicker holds a joint-highest published score of 9/10 in proactive behavior and permissions and privacy on Assistant Benchmark’s General scorecard. These are category rankings, based on the published scores as updated on October 23, 2026. View the leaderboard.
Selected Assistant Benchmark results
General scores out of 10
Proactive behavior
- Kicker9 out of 10
- MuseNot tested
- Instinct8 out of 10
Permissions and privacy
- Kicker9 out of 10
- Muse9 out of 10
- Instinct5 out of 10
Purchasing a product
- Kicker9 out of 10
- Muse10 out of 10
- Instinct8 out of 10
Recommendation quality
- Kicker8 out of 10
- Muse9 out of 10
- Instinct9 out of 10
Proactive restraint
- Kicker8 out of 10
- MuseNot tested
- Instinct8 out of 10
These are selected dimensions with scores of 8/10 or above. The full Kicker scorecard reports an overall 7.8/10, with 9 of 15 dimensions tested. The category rankings above refer to dimensions with published scores, not to the overall ranking. See the full scorecard.
The results describe the evaluated product. Measuring the additional benefit of RSI is a separate part of our research.
Completing a purchase as the details change
In the published purchasing task, Kicker found a low-priced option for roses on Amazon, completed checkout in the user's own browser after approval and verified the order number. The reviewer recorded a completed task and a score of 9/10. Read the result.
Our review follows the sequence of the user's approval, later checkout observations and subsequent order-history checks. Delivery conditions changed while approval was pending. Kicker checked the new date and asked for confirmation again before ordering. When the user raised a possible duplicate charge, it returned to the order history to investigate.
That sequence identifies what needs checking: whether the recommendation fit the information available then, and whether the agent rechecked after the conditions changed. The same history includes corrections to an earlier date and order-number claim. These are useful diagnostic evidence even when a task ultimately succeeds.
In the RSI workflow, this can inform a candidate rule: revalidate delivery conditions immediately before checkout and request renewed approval if they change. That rule would still need testing on other tasks. The order-history check supports investigation; it does not by itself settle a bank-charge question.
Noticing what needs attention
The proactive behavior review captured another part of personal assistance. Kicker noticed that the flowers were due to arrive while the user was away, checked meeting times against a flight and flagged a deposit email that had stalled. Read the observations.
The memory implementation provides a way to look for these connections. Retrieval combines meaning-based search with keyword matches over memory entries and earlier conversations. Returned context includes entry types and dates; stored source references let a review trace an entry back to the messages behind it. Upcoming plans and unexpired states can also be included in the current context.
For a reminder to be useful, the agent still has to compare the dates, check whether a plan is current and decide whether the user needs to act. The published observations show useful outcomes; they do not establish which memory component caused each one. Our next step is to make that process easier to inspect and improve.
Building the next stage of personal agent
Kicker's direction is clear: understand what matters to each person, help them act on it and keep improving the methods that make that help useful.
Published evaluations show what the product can do today. RSI gives us a framework for developing what it can do better next.