Ex-Googler Open-Sources Agent Memory System with 9 Context Tricks

A self-described 61-year-old former Google engineer has open-sourced a GitHub repo on context management to fix Agent memory problems. According to the author, the system has been quietly adopted by 3 frontier labs, with no accompanying blog post or launch event. He later shared 9 context management tips distilled from 3 weeks of use, relayed as a thread series by @bigaiguy, with several already published.

Confirmed

• The tips series was relayed point by point by @bigaiguy: tip one is the opening content; tip two is "relevance truncation"—most people keep full conversation history just in case, but every irrelevant token dilutes the important information, so irrelevant messages should be summarized and discarded; tips four and five both revolve around the "state block."
• The "state block" means that before each complex step, the agent writes down its state in three lines—what I know, what I'm still missing, what I'll do next—so it re-anchors itself every turn instead of relying on a decaying context window.
• Tip five also covers techniques like anchor re-injection and a memory ledger, rounding out the series; as noted above, "structured beats prose," meaning memory should not be stored as paragraphs.

Unconfirmed

• The claim of being "quietly used by 3 frontier labs" comes from the author; the posts relayed by @bigaiguy give no lab names or other evidence.

Why It Matters

• The series targets a widespread pain point in long-running Agent tasks: a gradually blurring context window and diluted memory. The techniques offered (state blocks, relevance truncation, structured memory) are low-cost to implement and immediately applicable, giving direct reference value to developers building Agents.

2026-09-14 ~ 2026-09-14 · 5 related posts

• [source] Agent context tricks: state blocks and anchor reinjection against long-task drift — bigaiguy · 2026-09-14
• [source] Ex-Google engineer open-sources agent memory system; 9 context tricks detailed — bigaiguy · 2026-09-14
• [source] Agent context trick 2: the Relevance Cutoff — summarize and drop irrelevant messages — bigaiguy · 2026-09-14
• Agent context trick 4: write a three-line state block before each complex step — bigaiguy · 2026-09-14
• The State Block: three lines of self-summaries before every agent step — bigaiguy · 2026-09-14