Memory agent founder: two-thirds of small-model memory failures happen with evidence in context
Archilas-Memory · reddit · 2026-10-12
The solo founder of Archilas, a memory layer for AI agents, shares what he learned: notes (exact values, names, dates, superseded facts) are compacted into a KV cache rather than a vector DB, so models answer from loaded notes with sources. On a long-conversation memory benchmark they hit 72%, but roughly two-thirds of misses occur even when the correct evidence is already in context — the small model hedges, abstains, or picks a similar event from another session. Retrieval tweaks, LoRA fine-tuning and note rewrites didn't fix it; a bigger model fixed more answers but started answering unanswerable questions. MCP/API/SDK access coming soon.
More from coding & agent
- Ex-Microsoft dev says AI could save Windows; low test coverage still sinks AI coding — julianharris · 2026-10-12
- Agent-drawn Excalidraw slides went from terrible to perfect in one year — HamelHusain · 2026-10-12
- Self-organizing agent teams hit 66.7% on math benchmarks vs 48.8% for their strongest member — zainhas · 2026-10-12
- Practitioner's verdict on autoresearch: great at speeding experiments, not at frontier runs — iaindunning · 2026-10-12
- levelsio cancels nearly all SaaS subscriptions: money now goes only to AI inference, content, data centers, power and taxes — karlwaldman · 2026-10-12
- Grok Bot is winning the agentic personal assistant race, says dev — Arindam_1729 · 2026-10-12