Alexandria¶
Give your AI coding agent a persistent, citable knowledge base. alexandria ingests papers, articles, code, and conversations into a local wiki with verified quote citations — and exposes it over MCP so Claude Code, Cursor, Codex, and other MCP-capable agents can retroactively query, cross-reference, and synthesize everything you've ever fed it.
pip install alexandria-wiki && alxia init
alxia ingest https://arxiv.org/abs/2307.03172
alxia mcp install claude-code
# now ask your agent — the knowledge is there, with citations back to source

Live belief graph from a real vault — 668 nodes, 500 edges. Blue =
code-derived, yellow = topic hubs, orange = web, green = papers.
Generate your own with alxia export <dir> --format graph.
Three things alexandria does that most "feed-your-LLM-a-folder" tools don't¶
- Durable across sessions. Knowledge accumulates in
~/.alexandria/; every ingest is a staged, verified write, not a regenerate-from-scratch pass. - Every claim has a citation. A deterministic verifier checks that each footnote maps to a verbatim quote in the original source via SHA-256 anchors — no hallucinated references.
- Belief revision, not just retrieval. Structured claims with
supersession chains. Ask
alxia why <topic>and you get what was believed, what superseded it, and why.
Capabilities¶
- Ingest from 14+ source types: files, directories, URLs, git repos, PDFs, code, RSS, IMAP, YouTube, Notion, HuggingFace, Obsidian vaults, archives, Claude Code conversations
- Search with hybrid BM25 + recency + belief support scoring
- Query with LLM-powered grounded answers that cite sources
- Track beliefs as structured claims with supersession chains
- Extract code structure from Python, TypeScript, Rust, Go, Terraform, Ansible, YAML via tree-sitter / AST
- Capture conversations from Claude Code sessions with artifact extraction
- Cascade merge — new sources on existing topics merge into the appropriate page with per-source attributable sections
- Cross-reference related wiki pages with auto-discovered See Also links
Core design¶
- No vectors, no RAG. The agent IS the retriever; alexandria exposes FTS5 + grep + belief graph + cross-ref follow as primitives.
- Filesystem is source of truth. SQLite is a materialized view;
lose the DB, rebuild via
alxia reindex --rebuild-beliefs. - Every wiki write goes through a deterministic verifier. Commits fail if citations don't validate.
- Hybrid search. FTS5 BM25 + recency decay + belief support as a composite score.
One-line capability check¶
alxia bench
Sample output from a real vault:
864 pages · 15,589 beliefs across 3,804 topics · 1,664 verified ingests
· search 41ms / 386ms P95 · citation-verified rate: 99.9%
Learn more¶
- Getting Started — full user guide
- CLI Reference — all commands
- MCP integration — scope, workspace pinning, gotchas
- Mobile vault via GitHub — read on any device, capture on the go
- Architecture — design documents