agent-lit 0.5.0

Persistent memory for AI agents doing literature review with your taste.

(-o-) alit

Your agent forgets every paper it's ever read.

alit is a local knowledge base for AI agents doing literature review. One agent reads 50 papers and stores structured summaries, citations, and reading status. The next agent — or the same agent next week — queries that knowledge instantly.

The agent builds the knowledge. You set the taste. Knowledge compounds.

pip install agent-lit → zero dependencies, SQLite-only, works with any coding agent.

The core loop

alit taste "vision-language grounding, embodied AI"   # you set direction
alit recommend 5                                       # agent picks what to read
alit summarize <id> --l4 "..." --model claude          # agent stores findings
alit ask "what approaches exist for X?" --depth 2      # agent synthesizes

Run alit --help for the full command list (25 commands for search, import, export, citations, and more).

What alit does and doesn't do

	alit	The agent
Stores papers, summaries, citations, taste	✓
Ranks recommendations (PageRank + taste + recency)	✓
Retrieves context for synthesis	✓
Persists across sessions	✓
Reads papers		✓
Writes summaries		✓
Decides what to cite		✓
Answers research questions		✓

alit stores and retrieves. The agent thinks. You set the taste.

Setup

pip install agent-lit    # or: uv add agent-lit

Spin up Claude Code, opencode, Cursor, or whatever you use, then prompt:

Set up alit for literature review in this project. See https://github.com/Zhou-Hangyu/alit

From there:

Find and add the top 10 papers on vision-language grounding from the last 2 years.

Read the next 5 recommended papers and summarize each one.

What does the literature say about cross-modal attention mechanisms?

How it works

.alit/
├── papers.db    ← one SQLite file, entire knowledge base
└── pdfs/        ← auto-downloaded from arXiv

No servers. No API keys. No vector databases.

Update

pip install --upgrade agent-lit

Under the hood

• Search: BM25 via SQLite FTS5
• Ranking: PageRank on citation graph (pure Python)
• Recommendations: PageRank + recency + taste matching
• Synthesis: multi-stage funnel retrieval (~5K tokens to query 10K papers)
• Enrichment: arXiv API (batched) with Semantic Scholar fallback
• Backward compatible: schema auto-migrates on upgrade

License

MIT