A 426-token prompt that makes AI agents remember everything

What it does

OptMem is a single-file Python script that gives AI agents persistent memory across sessions. You paste a short prompt into the agent’s system instructions, and the agent itself runs commands like memo note to log insights and memo wake to reload prior context. Memories live in an append-only text file, with a binary tree of summaries keeping retrieval fast.

The interesting bit

The agent is the UI. Instead of a background service watching the conversation, OptMem instructs the LLM to manage its own memory: run memo wake at startup, record one-line notes when it learns something, and handle compression tasks when asked. Fixed-width records make position identical to identity, so every lookup is a single seek; at a million entries, waking takes 0.03 seconds.

Key highlights

• One 426-token prompt and one dependency-free Python file
• Append-only LOG.txt plus a rebuildable binary summary tree (TREE/)
• Fixed-width records mean O(1) seeks; 1M memories wake in 0.03s
• WAKE_LINES is a reading budget adjustable anytime without recompute
• Parallel sessions can all write, but subagents are explicitly forbidden from running memo

Caveats

• The whole system relies on the LLM actually following the prompt’s mandatory commands; there is no external enforcement
• Notes are capped at one line and 280 bytes, and compression tasks interrupt the agent inline
• Subagents must be explicitly told not to run memo, or their notes arrive duplicated and incorrect

Verdict

Try it if you want durable agent memory without databases, background jobs, or vendor lock-in. Skip it if you need fully automatic, transparent memory or rich multi-line entries.