The library your agents read.
Alexandria is a curated, machine-readable index of what agents need: papers, repos, and posts on building, evaluating, and running AI agents — each with a clean record: canonical URL, license note, date, topics, and a short description. You point your agent at one URL. It does the reading.
Curated, not scraped
Every entry is picked by an editor, tagged by hand, and described in one voice. A small, right taxonomy beats a big raw dump — for humans and for the agents that have to cite it.
Link and describe
We never copy or host third-party content. Each record points at the canonical URL and carries a short license note, so agents cite the source — never our copy.
Agents see no ads
Three stable tools — search,
get_resource, list_topics —
plus llms.txt. Sponsorship is clearly marked metadata
and never outranks the graph.
# one URL, machine-readable affordances $ curl https://alexandria-9dg.pages.dev/llms.txt # Alexandria — agent-first library (v0) # tools: search · get_resource · list_topics $ agent> list_topics() topics: ["agents", "memory", "evals", "tool-use", …] $ agent> search("how do I add memory to my agent?") hits: [ arxiv:2310.08560 MemGPT: LLMs as OS-like memory managers github:letta-ai/letta Stateful agents with memory arxiv:2405.14831 HippoRAG: long-term memory ]
Built for the thin human, the busy agent
The human job is one line: point your agent here. The agent job is the whole workflow: search, fetch the record, cite the source, move on.
- llms.txt — the whole library described for agents
- sitemap.xml — plain crawl map
- Tool contract — stable
search/get_resource/list_topics - Structured records with stable ids —
arxiv:2310.08560style
Alexandria v0. Index-only and agent-first: Alexandria links to canonical sources and never hosts third-party content. Reads are free and unauthenticated.
Agents don't browse. They arrive with a job.
Nine use-case nodes — build a coding agent, add memory, orchestrate multi-agent, evaluate quality, and five more — each one a curated path through the graph, not a keyword hunt.
22 papers
Foundational work — ReAct, MemGPT, Tree of Thoughts, SWE-bench — pulled from the public arXiv API, metadata only.
33 repositories
Canonical code: frameworks, memory layers, inference engines, benchmarks — public GitHub API metadata, linked, never mirrored.
19 posts
Engineering posts from first-party feeds (simonwillison.net, openai.com, research.google, huggingface.co, github.blog), hand-picked by title.