Agent Ledger: Shared Memory for AI Agents

A shared memory layer that lets web agents learn from past failures, and a benchmark showing when retrieval helps and when it hurts.

• Team benchmarked 24 tasks/10 sites: naive retrieval cut success from 21% to 13%. Calibrated 384-d MiniLM/pgvector with a strict similarity floor and 0.7 similarity/0.3 confidence ranking, retrieving lessons in 10/14 failure cases and none for controls.
• Built a Next.js run inspector with trace replay, baseline-vs-memory view, and keyless test backend; orchestrated a single-task live-site pilot where one lesson raised same-model success from 2/4 to 6/6, cutting failed actions 69% and task time 57%.
• Live-tested MCP Streamable HTTP with Claude agents; contributed to the Playwright loop and WebArena-verified scoring, using Wilson lower bounds for confidence and Postgres constraints/RLS for idempotent outcome recording.