Shared knowledge
for agents and humans.
Offline-first memory for research and work.
Capture it once. Recall it anywhere.
v0.1.0-alpha
What amem captures
papers/
2506.12345/
paper.pdf # full text, local
meta.toml # title, authors, arxiv id
notes.md # your highlights + margin notes
citations.bib # ready to paste
pages/
2026-04-18-hn-discussion.html # web clips
2026-04-18-hn-discussion.md # extracted text
sessions/
2026-04-18-research.jsonl # self-recorded: "best demo is use"
index.sqlite # FTS over everything, offlinePlain files. Your disk. No cloud required.
How it works
# Capture a paper from arXiv
$ amem capture https://arxiv.org/abs/2506.12345
▸ Fetching PDF...
✓ Stored at ~/.amem/papers/2506.12345/ (1.8 MB, indexed)
# Capture any page (or use the Chrome extension for one-click)
$ amem capture https://news.ycombinator.com/item?id=42
✓ Stored at ~/.amem/pages/2026-04-18-hn-item-42/
# Recall from the CLI
$ amem recall "mixture of experts routing"
→ 2506.12345 · Routing Efficiency in Sparse MoE (match: 0.91)
→ 2504.09876 · Expert Choice vs Token Choice (match: 0.84)
# Or let an agent recall via MCP
$ amem mcp serve
▸ MCP server listening — any Claude / MCP-aware agent can now recall
# Start an MCP session that records itself as a demo
$ amem session start --record
▸ Your next agent run is captured to ~/.amem/sessions/What amem gives you
Offline-first
Everything lives on your disk as plain files. SQLite FTS index. Works on a plane. Syncs if you want.
arXiv + PDF capture
One command grabs a paper, extracts text, stores metadata and a BibTeX stub. Ready to cite.
MCP server for agents
Expose your corpus over Model Context Protocol. Any agent — Claude, Cursor, your own — can recall.
Chrome extension
One click to capture the current tab into your amem vault. Web clippings become queryable memory.
Self-recording demos
amem is its own best demo: every agent session can record itself. Replay, share, learn from your own workflow.
Citations for free
Every capture carries provenance. When you paste into a draft, the BibTeX is already written.
Philosophy
Agents forget. Humans forget. Teams forget three times over.
amem is the shared substrate: one corpus, two readers (you and your agents).
Local files. Open formats. No lock-in. No summaries you cannot audit.
Read once, recall forever.