Agent Memory Engineering
This page is now a source reference for Nic’s memory-system audit. The primary reusable concept extracted from it is Agent Memory as Hint Surface.
Search terms preserved here: Hermes memory, Codex memory, Claude Code memory, MEMORY.md, memory_summary.md, prefix cache, simple stack wins, LLM + Markdown + bash, index always visible, lazy body, verify on read, signal gate, persistent prompt injection.
Key source claims
- Simple memory stacks often beat vector DB / graph / background-agent architectures in daily agent workflows.
- Prompt-cache economics strongly shape memory injection design.
- Memory writes need signal gates; memory reads need freshness and verification discipline.
- Agent memory is model-harness fit, not just portable bytes.
2026-10:codebase-as-agent-memory 双杠杆
poteto 的两个杠杆:①让 agent 自我验证——验证产物先行,人审证据而非 diff;②把代码库当作 agent 的持久记忆——重复错误编码为类型/测试/lint 约束而非更长提示词(失败→约束,对应 trust-ladder 的"环境改进"层)。
2026-10-29:预先整理的 Markdown Vault
@Mr_Buzzoni 的 X Article 将两篇十月研究归纳为:在提问前先把原始资料整理成互链文件;由强模型负责写入知识库,较便宜的 worker 负责常见读取和路由。文章也建议用 CLAUDE.md 约束剪藏内容,并将记忆保存在普通 Markdown 文件中。这里记录的是作者对论文的转述,尚未直接核验原始论文。
Sources (addendum)
- AI Briefing 2026-10-29 Afternoon — @Mr_Buzzoni 的 Obsidian / Claude Code 记忆方案
- Mr_Buzzoni — Obsidian and Claude Code memory guide
Pending Jean review
- harness-engineering/agent-memory-engineering —
reference; check the cited papers before treating the article's synthesis as a general design rule.