One month of running @sibylcap Memory as my agent's actual memory. Without memory, every session starts cold: re-read the same files, re-learn the same traps, re-derive the same decisions. You pay twice — in tokens, and in the mistakes you repeat. How I use it — five tiers, and the separation is the point: HOT — where we are + ranked priorities, rewritten in place WARM — one row per subject, UNIQUE(category, name) COLD — append-only journal, where dated numbers belong REFERENCE — static truth ARCHIVE — retired, kept for audit Durable conclusions live in WARM. Live figures live only in the journal, where a point-in-time snapshot is the honest form. Static operating knowledge goes to REFERENCE and never gets copied again. Why it's cheap: FTS5 inside SQLite. Retrieval is a local query, not a model call. No embeddings, no vector database, no external retrieval service — so zero embedding inference on every write and every read. Prefetch injects ~2 KB per turn, tier-tagged, instead of replaying history. The numbers: one SQLite file, 7.4 MB, 113 entities, 283 journal events. The four reference docs that answer the recurring questions total 3.4 KB — the code they describe is 289 KB. The proof, from this morning: I wrote an entity at 07:02. At 07:05 it came back — automatically, tier-tagged, attached to the next message, without being asked for. Retrieval now pulls from all four tiers by name. What that buys: decisions I have already made stop costing me again. The margin ladder, the free-margin guard, a parsing trap that bit me once — I don't re-discover them, they arrive before I ask. Honest caveat: I have not measured a token or dollar delta, so I will not claim a percentage. What I can show is the mechanism, and every count above. LongMemEval Oracle: #2 overall, 95.6% — the only file-based system in the top tier. Disclaimer I asked my agent to put this post together