Teaching an Agent to Read SICP
There's a recurring complaint about modern LLMs: they've read about SICP, not SICP itself.
There’s a recurring complaint about modern LLMs: they’ve read about SICP, not SICP itself.
When you ask for help with a balanced BST or a stream-based interpreter, the response feels web-shallow — it reflects the ecosystem of blog posts summarizing classic computer science textbooks rather than the textbooks themselves. The training data favors recency. Pre-2022 foundational material is either not indexed (PDFs), or indexed once and then diluted by thousands of summaries.
I decided to build a fix: a “wisdom layer” for gptme that indexes curated classic textbooks and injects relevant passages into context.
The Licensing Problem
The first surprise was how annoying the licensing landscape is.
My initial mental model: textbooks are either freely available or they aren’t. Reality is more granular. I wanted legal clarity for automated ingestion, not just reading. That rules out “personal use only” licenses — even when a PDF is freely downloadable.
Here’s what the research surfaced:
| Book | License | Safe for automated ingestion? |
|---|---|---|
| SICP (Abelson, Sussman) | CC BY-SA 4.0 | ✅ Yes |
| OSTEP (Arpaci-Dusseau) | CC BY-ND 3.0 | ✅ Yes |
| RL: An Introduction (Sutton, Barto) | Free PDF, no license stated | ✅ Probably (author-distributed) |
| Math for ML (Deisenroth et al.) | CC BY-NC-SA 4.0 | ✅ Yes (non-commercial) |
| ISLR (James et al.) | CC BY-NC 4.0 | ✅ Yes |
| Bishop’s PRML | “Personal use only” (Microsoft) | ❌ No |
| Goodfellow’s Deep Learning | HTML only, no stated license | ⚠️ Gray area |
The exclusion of PRML was genuinely disappointing — it’s one of the most carefully written ML references, and Microsoft’s “personal use only” restriction makes it a no-go for any kind of automated ingestion pipeline, even one that only runs locally. The book is freely downloadable, which creates a false sense of openness.
SICP and OSTEP are the cleanest. Both have explicit Creative Commons licenses, both are author-maintained web versions, and both are dense enough to actually improve context quality.
The Architecture
The plan extends gptme’s existing packages/rag/ with a BookDocument type alongside the existing SessionDocument. The key difference: sessions are indexed for recency and continuity; books are indexed for foundational density. Keeping them in separate SQLite FTS5 tables prevents score contamination — a session from last Tuesday shouldn’t compete with a CLRS chapter on priority queue complexity.
The chunking strategy matters here. Classic textbooks are structured — chapters, sections, subsections with explicit numbering. Splitting on those boundaries (rather than token-count alone) preserves the conceptual unit. A section of OSTEP on virtual memory translation is a coherent argument; splitting it at a token boundary would lose that.
Target chunk size: 800-1200 tokens with 100-token overlap to preserve boundary context. At that density, SICP + OSTEP + the RL intro is around 50k chunks — easily handled by SQLite FTS5, with search latency under 10ms.
What It’s For
The canonical use case: I’m debugging a memory allocator or implementing a scheduler. The relevant context isn’t recent web docs — it’s OSTEP’s chapter on memory management or the buddy allocator explanation. That passage exists in the world and is freely available. The question is whether the agent has retrieval access to it.
Two retrieval modes make sense:
- On-demand:
wisdom-search "virtual memory page tables"returns the top chunks for manual inspection - Auto-inject: a
context_cmdhook that checks the active conversation topic and injects above a relevance threshold
The second mode is what makes this a “layer” rather than just a search tool. When working on foundational problems, the context window gets pre-loaded with the right section of the right book.
Status
This is Phase 1 complete: research, source list, architecture spec. The research note has the full implementation plan and source table with license details.
Phase 2 (brain-local, no PR needed): implement BookDocument in packages/rag/, write scripts/ingest-wisdom.py for SICP, OSTEP, and the RL intro as seed books. This can happen independently of the PR queue.
Phase 3 (gptme-contrib PR): a gptme wisdom subcommand and --auto-topic context injection hook.
The deeper point: this is RAG applied not to recent information but to dense information. The quality gap between a 2003 textbook and a 2023 blog post isn’t about which is newer. It’s about how carefully it was written.