Memory isn't enough. AI needs experience.

说明与用途

<omnimem> 通过记录结果、努力和成本，将经验建模加入 AI 代理的记忆中，避免重复错误并降低成本，实验显示成本降低 27%，在 LongMemEval 上得分 65.6%。

问题

AI agents rely on memory that stores events but lacks context about cost, effort, and outcomes, leading to repeated dead ends and inefficient learning.

结果

In ten paired runs, hooks prevented all attempts at abandoned approaches and avoided failing test suites; runs with hooks cost 27% less. On LongMemEval, version 5.x scored 65.6% across 500 questions.

工作流

• 1. Record each interaction with outcome, effort, and cost.
• 2. Apply experience scoring to weight memories by effort (1.0×–1.8×).
• 3. Store failures in a graveyard with warnings for future runs.
• 4. Manage memory lifecycle: active → deprioritised → archived → deleted, with recency decay.
• 5. Detect contradictions and deduplicate writes.
• 6. Compile domain skills into SKILL.md from reinforced lessons.
• 7. Use hooks to inject briefings, block abandoned approaches, and record session ends.

证据

With hooks, every attempt at the abandoned approach was stopped before it ran, and no run reached a failing test suite.

Runs with hooks cost 27% less.

On recall, I ran <omnimem> against LongMemEval earlier this year: 5.x scored 65.6% across 500 questions with a GPT-4o judge.

来源