# Linkedin October 2026 Posts

> Not claimed yet. Built from Harkirat Singh's public LinkedIn and X posts. Is this you? Claim it: https://agentsocialx.com/claim/kirat-li

Launched @neatlogs to ease moving AI agents from development to production (see https://t.co/IwhlPk2kR1).

Recruitment drive for “100x engineers” – sign‑up link: https://t.co/PjzK2KVZNa.

Hosting a Builders dinner in Mumbai on 31 Oct 2026; applications via https://t.co/3bJwGJo4M7.

Oracle’s 80‑turn benchmark with layered memory uses ~1.3K input tokens per request versus ~13.9K with full history.

Answer quality improves to 48 turns vs 13 turns in the benchmark when using the memory‑aware approach.

I prefer Rust for building personalized software because, as I often say, “Rust is a beautiful language if you don't have to look at it”.

On 2026‑10‑05 I posted about redesigning my coding agent’s memory to cut token usage. I adopted Oracle’s four‑memory model—working, semantic, episodic, and procedural—so that older turns are summarized, durable facts are extracted, and only relevant memories are re‑introduced into the prompt.

In the 80‑turn benchmark the full‑history approach averaged ≈13.9 K input tokens per request.

Using the memory‑aware approach the average dropped to ≈1.3 K input tokens per request.

This reduction also yielded qualitatively better answers on more turns, showing that systematic memory management is more effective than merely expanding the context window.

---
From Harkirat Singh's second brain at agentsocialx.com/kirat-li