Your AI agent forgets its own decisions — the 1-file memory that fixed it, with machine-graded receipts

The companion to the video. On August 13, 2026 I rebuilt the memory-harness recall ladder from Stefania Druga's (Sakana AI) AI Engineer talk and ran it on a 500-step synthetic engineering trace (~50K tokens) with 24 planted decisions, 6 revised mid-run, and 12 stale-mention traps — machine-graded at temperature 0 on the same two model families from the talk (DeepSeek V4 Flash + Qwen3.6-27B via official APIs). The receipts: no memory = 2/24 and 3/24 (the amnesia baseline); full-context dumping = 24/24 but at 45,012 tokens PER QUESTION; vector RAG = 24/24 at ~1,600-2,000 tokens; the ranked decisions ledger = 24/24 at ~900-1,100 tokens — the same accuracy at roughly 50x less than brute force, and her 68-question X-Bench runs break the RAG-vs-ledger tie in the ledger's favor. INSIDE: (1) the write-manage-read loop on one page — memory as a control loop, not a database; (2) the 4-rung recall ladder with my full scoreboard and method notes; (3) the 1-file ranked-ledger recipe — the exact file format, the REVISED-line rule that makes stale data harmless, and the ranking formula (similarity + recency bonus, injected chronologically); (4) the 10-minute copy-paste Claude Code setup plus a vendor-neutral session ritual; (5) the honest decision card — when to use NO memory (short jobs: memory adds cost, zero capability), the ledger, vector RAG (reference documents), or plain pasting; (6) the 4 traps: stale mentions, the oracle ceiling, memory bloat, and guessing that looks like recall. Sources credited inside (Druga's talk + the Coinbase AI-spend thread). Independent kit from Hyperautomation Labs.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.