Adam Selene
Give an agent memory that travels with it when the model changes. Explore the architecture and the limits of its evaluation.
Explore the memory system →Read the evaluationIndependent research · Airway Heights, Washington
AI agents that remember.
Evidence of what works.
Persistent memory. Practical inference. Reliable agent systems. Boundary Labs turns experiments on accessible hardware into results you can inspect and work you can build on.
Research you can use
Three ways into the work.
Start with a system, a research question, or a practical problem.
Inference research
Find the tradeoffs between speed, tool use, and quality across consumer GPUs, Apple Silicon, and an older laptop CPU.
Browse the results →Read the evaluation methodContext Farm
Make the rules and exceptions in one person's head usable by a team. See a worked dispatch example with a source trail.
Try the example →Explore the prototypeEvidence, with its limits
Results worth looking into.
Recall with memory supplied
88% on a 25-question single-session-user LongMemEval subset, using context-window injection. Not the full benchmark.
Read the study and limitations →Median CPU task latency
Qwen3-30B-A3B, 15-task agent-fitness evaluation, 2026-08-24. Single-shot result; the follow-up multi-turn workflow was not viable.
Inspect the CPU experiment →Logged benchmark runs
Historical corpus as of 2026-06-27. A count of runs, not independent replications or a claim about today's activity.
Browse the public artifacts →From the research record
Recent investigations.
Old laptops. Useful work.
Omarchy, local dictation, a custom dock, and the rough edges that shaped the recent laptop work.
Can an older laptop run a useful local agent?
MoE and dense models compared on tool discipline and task latency. Includes the multi-turn failure.
Evaluate the decision, not just the answer.
Practitioner-judged scenarios and the boundary between public results and held-back evaluation items.
Faster inference can still be the wrong choice.
The speed and quality tradeoff in the consumer-GPU MoE campaign.
The person behind the lab
One operator.
A working research practice.
I'm Dino Vitale, a systems architect in Airway Heights, Washington. I run Boundary Labs to investigate how AI agents retain useful knowledge, how deployment choices affect their behavior, and what it takes to keep them useful over time.
The work starts with an operational question and ends with evidence: an experiment, an implementation, a failure analysis, or a published research artifact.
Read my writing → Papers and preprintsLab status · verified 2026-09-06
Local operations. Swappable inference.
Production backend: GPT-5.6 Luna (OpenAI API)
Verified 2026-09-06. Agent orchestration and memory run locally; production inference uses an external API through a shared gateway. This is a dated verification, not a live status feed.
Consumer-GPU experiments remain available as dated research. The dedicated GPU tower was retired on July 20, 2026; CPU research continues on commodity hardware.
See the current setup and hardware history →Work with Boundary Labs
Bring a question we can test.
Research collaboration, scoped consulting, and hardware or compute sponsorship tied to a concrete deliverable.
Explore collaboration proposals →[email protected]