neoneye/agent-memory-atlas：AI 智能体记忆系统实地指南
neoneye/agent-memory-atlas
neoneye/agent-memory-atlas 发布，一份面向 AI 智能体记忆系统的代码级实地指南，对比各系统的记忆单元、存储模型、写入路径、检索机制、纠错语义、信任模型与运维风险，每份报告均链接到所分析的确切 Git commit。项目覆盖源码公开可读的系统，其中十六个非开源系统被点名而非取平均值处理，并附可复用的记忆架构模式指南与跨系统对比分析。
A code-grounded field guide to memory systems for AI agents.
The atlas compares memory units, storage models, write paths, retrieval mechanics, correction semantics, trust models, agent integrations, and operational risks across systems whose source is publicly readable — most open source, sixteen of them not, which the atlas names rather than averages over. Every report links to the exact Git commit that was analyzed.
Read the atlas
Visit neoneye.github.io/agent-memory-atlas.
Discuss
Questions, corrections, and systems worth adding are welcome on Discord: neoneye.github.io/agent-memory-atlas/discord.html. That page redirects to the invite, so the invite code stays in one file and can be rotated without editing every place it has been shared.
Repository structure
content/systems/— individual repository reports.content/patterns/— reusable memory architecture pattern guides.content/overview.md— cross-system comparative analysis, rendered to/overview/.content/families.md,content/compare.md,content/appendix.md— the system families, the generated matrix, and the pins, licences and known limitations.content/methodology/— the review and synthesis formats.site/— the designed homepage and the/discord.htmlinvite redirect.templates/— the shared document-page template.assets/— styles, behavior, and social-preview media.docs/— generated static site published by GitHub Pages.AGENTS.md— the entry point for an AI agent: what to read to build memory for a product, in what order, and what not to read..agents/protocol/— the machine-readable half: portable acceptance tests with stable ids, and the build-brief, closure-report and lock-file formats..agents/skills/use-the-atlas/— designing memory for another product from the atlas: profile, brief, approval, then code..agents/skills/screen-repository/— the security screen to run on a checkout before it is read or run, required for new and re-pinned work. What has actually been screened, and against which pin, is recorded innotes/screening/screening.jsonand reported byscripts/check_screening_ledger.py; a large share of the older corpus predates the ledger and is unscreened..agents/skills/add-memory-system/— the repeatable workflow for researching and integrating another memory system.
Contribute
CONTRIBUTING.md covers the three useful contributions — correcting a report, proposing a system, arguing with a pattern — and what the build refuses. The reader-facing version is neoneye.github.io/agent-memory-atlas/contributing/.
Add a memory system
Screen the checkout first — python3 scripts/screen_repo.py <path> reports auto-executing hooks, build-time execution and unpinned dependency surfaces without running anything from the tree. Then invoke the repository-local $add-memory-system skill with the path to a source checkout. It pins the analyzed commit, scaffolds the report, guides the code review, updates applicable design patterns, integrates the comparison and homepage, and validates the generated site.
Build locally
The build requires Pandoc.
npm run build npm run serve
Then open http://localhost:4173.
License
MIT
来源：GitHub · github.com