Active work across three areas: open infrastructure for personal devices, memory systems for long-running agents, and small models targeted at consumer-grade hardware. Notes, preprints, and code are linked below.
airdress/ ├── protocol-v2.md ⬤ ├── relay/ MIT ├── sdk/ Apache-2.0 └── specs/
Protocols, runtimes, and reference implementations for self-hosted personal-software stacks. Permissive licenses; specifications maintained alongside code.
Decay, reinforcement, and salience-weighted retrieval as alternatives to monotonically growing vector stores. Evaluating recall, latency, and storage trade-offs over multi-month traces.
Quantization, distillation, and runtime work targeting NPUs and SBCs in the 5–15 W envelope typical of home gateways and sensors.
Specifications, datasets, and reference code are released openly when the work is generalizable beyond the product. Closed work is identified as such.
Long-running, single-user agents have structural similarities to memory and attention in biological systems. We treat that as a useful prior, not a guarantee.
Model, runtime, and silicon are evaluated jointly. Reported results include power, memory, and latency on commodity edge hardware — not only accuracy.
airdress/ ├─ core/ │ ├─ relay.go │ └─ protocol-v2.md ├─ chat/ · MIT └─ edge-runtime/ · Apache-2.0
Most of what we build is open source — both the components that ship as part of Airdress and standalone tools that don't have a commercial home but matter to the wider ecosystem. We treat open source as the default, and write down our reasoning when something needs to be commercial.
Most agent designs put one large LLM at the center and bolt memory onto the side. We are working the other way: a tiered cognitive architecture — from a reflex layer through episodic, semantic, procedural, and parametric tiers — where the LLM is one participant the rest of the system invokes when lower tiers cannot handle a request, not the engine they serve. The interesting work is in the dynamics: consolidation that compiles repeated patterns downward, demotion when compiled skills misfire under environmental drift, and parallel cross-tier retrieval as a routing mechanism. Prior work covers parts of this — HippoRAG for hippocampal indexing, Titans for surprise-driven write gating, Larimar for CLS-style fast/slow consolidation — but no system integrates them with explicit dynamics. That is where we are pointed.
Personal software ends up running on small hardware — sensors, gateways, the box on your shelf. We are training and tuning models small enough to live on those devices, so that AI in the home is not a round-trip to a datacenter on every interaction.
Most of this work happens in the open. If you are a researcher, contributor, or company working on adjacent problems, we would like to hear from you.