You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
TaterLabs builds local-first AI infrastructure and studies the safety of AI agents that have persistent memory, embodiment, and autonomy. As agents gain long-lived memory and tool use, their failure modes shift in ways stateless evaluations miss: stale-memory reinforcement, manipulation via injected memories, hallucinated continuity, and unsafe goal fixation. Working solo on commodity hardware over the past year, I built Mnemosyne (a local, inspectable memory layer for LLM agents) and Nexus (an embodied evaluation environment), plus a published embodied-LLM/FlyWire experiment. This grant funds productionizing Mnemosyne into an open local memory API and releasing a reproducible benchmark for persistent-memory agent safety.
Within ~6 months: (1) release a production-ready local Mnemosyne memory API with import/export and documentation; (2) publish an open benchmark comparing stateless, vector-memory, and reinforcement/decay-memory agents; (3) document and demonstrate concrete failure modes (stale-memory reinforcement, memory-injection manipulation, hallucinated continuity, unsafe goal fixation) with mitigation patterns (decay, source attribution, confidence scoring, memory quarantine, audit logs); (4) publish reproducible code that runs cheaply on commodity hardware so others can replicate it.
Requesting $25,000 for ~6 months full-time: roughly 70% stipend (inclusive of tax), 15% compute/hardware/hosting for benchmark runs, 10% documentation and public release, 5% buffer. At the $5,000 minimum I would ship the core local Mnemosyne memory API and initial documentation; the full $25,000 covers the benchmark, failure-mode studies, and a reproducible public release.
Solo founder (TaterLabs LLC, Iowa); self-taught builder. Over the past ~year, solo and self-funded, I built TaterTOS64 (a GPL-3.0 64-bit operating system from scratch), Mnemosyne (a local memory server for LLM agents), Nexus (an embodied evaluation environment), and a published embodied-LLM/FlyWire connectome experiment (Zenodo, with code and video). Portfolio: github.com/legacyindiesubmissions-ai
Most likely failure: as a solo independent researcher without formal affiliation, the benchmark could end up less rigorous or less widely adopted than hoped, or the scope could exceed six months. Mitigation: the core memory-API release is independently useful even if the benchmark is partial, and I will scope tightly and publish incrementally. Worst case, the outputs are a working open local memory tool plus documented failure-mode observations - still useful public artifacts.
$0 in outside funding to date; the work has been self-funded. In June 2026 I also applied to the Survival and Flourishing Fund (including a $25k Speculation Grant request), Emergent Ventures, and the Long-Term Future Fund; all are pending.