PARM

PARM is a causal benchmark suite Karthik built for testing when an AI agent should use personal memory during a real workflow. It asks whether an agent can recognize when a late observation makes a specific memory actionable, retrieve it before the action it governs, and stay silent when the cue is removed. Prompt time retrieval cannot see a cue that only arrives in a late tool result, and naively searching the whole observation floods the agent with unrelated personal context. Key details: Every scenario is a causal triplet: a positive case where a late cue makes a memory relevant, a cue ablated control with the cue removed, and a memory included ceiling that confirms the intended action is achievable PARMBench v1 calibration batch: 157 frozen base scenarios, 471 executable cases, and 140 persona isolated histories; the final frozen replay passes 157 of 157 scenario contracts Retrieval agnostic: the benchmark supplies raw history and a causal contract, not a preferred memory store PARMBench Workflows, the primary track, scores multi step tool trajectories from the final environment state, timely memory admission, and intervention restraint, without an LLM judge. It currently validates 18 deterministic cases On the 54 case Amara development suite, PARM V5 made 30 of 36 correct positive and control decisions, against 18 of 36 for the strongest comparison conditions (no memory, all entity output RAG, and naive output RAG). PARM admitted 15 gold sources with 6 spurious admissions, while all entity output RAG admitted 16 gold sources with 1,361 spurious ones. This is a development result, not a held out claim Data, manifests, retrieval artifacts, and results are versioned so every claim can be replayed Apache 2.0 licensed