Category: Building Jarvis
Open research notes on building Jarvis — a persistent, self-improving AI agent. One paper at a time.
-
AEGIS: A Multi-Layered Security Framework for Autonomous AI Agents
The honest question isn’t whether your agent deployment could be a problem — it’s which specific problems apply to yours, and what mitigations are actually proportionate.…
-
Budget Prompting: Cutting the Cost of Always-On Memory Agents 2–3×
Leave your agent running overnight and the bill is brutal: every turn re-bills the whole context window. Budget Prompting is 20 techniques that cut per-turn cost…
-
PREFRONTAL: Giving Your Agent an Executive Function with a Recipe Execution Substrate
Your model is a brilliant worker and a terrible executive — PREFRONTAL is the missing executive layer: a recipe execution substrate that picks the right playbook,…
-
MNEMOSYNE: Four Hooks That Upgrade Your Agent’s Memory Without Forking It
Your long-running agent’s memory has four quiet failure modes — slow lookups, task-blind retrieval, silent contradictions, and detail destroyed at compaction — and MNEMOSYNE fixes all…
-
Why Your Pre-Push Privacy Gate Is Lying to You — and the Recipe as the Missing Middle Layer
Every individual push passed the privacy gate. The accumulated history still leaked 16 times — here’s the per-delta-vs-cumulative trap, and the named, replayable abstraction that sits…
-
Total Recall: Pointer-Based Compaction and Task-Conditioned Retrieval for Persistent LLM Agents
Most AI agents compact old context by summarizing it — quietly losing the one detail that mattered. Total Recall is a lossless, event-sourced memory architecture that…