concept · concepts/The Borges Map.md

The Borges Map

status: active · created 2026-11-08 · 2 sources · #method #agent-memory #knowledge-management #method-tension

A knowledge store that grows as large as the sources it summarizes "effectively abstracted nothing." It is a map at 1:1 scale: complete, and useless. Named for Borges's one-paragraph story "On Exactitude in Science," whose cartographers drew a map the size of the empire.

Evidence

Anthropic — The Components of Agent Memory, on an agent-built wiki for a legal benchmark: "On another LAB task, the wiki grew larger than the documents themselves … Claude had made an extremely detailed wiki that effectively abstracted nothing." The same study found that summarizing loses facts (41/57 survived) while verbatim quotes keep them (57/57). So compression should come from selection (what to keep) and structure (links, syntheses), not from paraphrase.

In this wiki (measured 2026-11-08, `scripts/lint/borges.py`)

Overall: wiki ≈ 0.70× all of raw/, or 0.92× of what had actually been ingested. That's close to the line.
The Borges cluster: engineering craft, 11.1×. A 632-word heuristics list grew into 11 single-source pages totalling ~7,000 words. The growth is the agent's elaboration, not the human's experience.
The Wiener excerpt: 9.8×. 122 words became ~1,200.
Healthy: most article-length sources compress to 0.2–0.8×.

The craft cluster is also T1 in Method Reflection. Willingham says maxims need worked examples to transfer (Metacognitive Strategies). The agent's paraphrase can't supply them, but the human's real cases can. One fix serves both: cut the elaboration and replace it with worked examples.

The reading-side cost

Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j) prices the same failure from the reader's seat: if the agent "consumes too much context every time it touches the knowledge base, both read and write slow down and the context window fills and rots faster." His alternative to lossy paraphrase is structural: "a summary can stay a pointer rather than a replacement, since the agent that needs the underlying detail can hop into the specific section or back to the raw source." That is the "keep what search can't re-find" rule with a mechanism: sections and links as addressable targets (Navigation vs Search).

The rule

Compress, don't restate. A page longer than its source is a red flag. A cluster longer than its source is a Borges map.
Keep what search can't re-find: connections, syntheses, tensions, provenance. Everything else stays in raw/.
Not every link adds structure. "The overlap must be the right overlap … A garbage can is full of overlap" (Alexander — A City Is Not a Tree).

Connections

Agent-Maintained Wiki — the pattern this failure mode belongs to.
Method Reflection — T5 (wiki as agent memory), and T1 (the craft cluster).
Working Heuristics — Data, Matching, Product, Org, Tooling — the main offender.
Tree vs Semilattice — the "right overlap" criterion for which links earn their place.

Sources

Anthropic — The Components of Agent Memory
Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j) — context rot; summaries as pointers.

Linked from 10

Agent-Maintained Wikiauto-ingested

A rough check would compare total wiki/ size with raw/ size.~~ Measured 2026-11-08: 0.70× overall, two clusters at 11.1× and 9.8× (The Borges Map).

Alexander — A City Is Not a Tree

The Borges Map — "the right overlap": not every link is structure.

Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j)

The Borges Map — reading-side cost of size; "a summary can stay a pointer" supports keep-what-search-can't-find.

Front Door Pattern Language

The Borges Map — a lens that shows everything is the 1:1 map; the Location lens and the todo are bounded for that reason.

Legibility in the Lens, Not the Land

The Borges Map — a lens that tries to show everything is a 1:1 map.

Method Reflection

And the Borges check measured the cost: the craft cluster is 11.1× its source.

Navigation vs Search

The Borges Map — a summary that stays a pointer to a section is compression without loss.

Pattern Language

That joins Willingham's critique (Metacognitive Strategies: maxims need worked examples) and the Borges finding (The Borges Map: the cluster runs 11× its source).

Tree vs Semilattice

One that just repeats a shared word is garbage-can overlap (The Borges Map).

Working Heuristics — Data, Matching, Product, Org, Tooling

[!warning] Borges + pattern-language check (2026-11-08): this list and the 11 pages citing only it total 11.1× the source's length (The Borges Map).

Links to 10

In the log 2

2026-11-08 ingest

Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j) (Zach Blumenfeld, Neo4j, 2026-08-31; agent-fetched at the human's request after discussion) — 1 source, 1 concept (Navigation vs Search: search vs following structure by pointer; the reverse index a file system lacks), 1 #to-read stub (Karpathy — LLM Wiki Gist, the primary source for this wiki's own pattern). Updated Agent-Maintained Wiki (scaling failure modes; 1 contradiction: "can't be scaled for personal workflows" vs Anthropic's 420M-token wiki; one open question answered by the Borges check, one added), The Borges Map (reading-side cost; summaries as pointers), Front Door Pattern Language (Location lens = Leiden communities; Hierarchy = in-degree), Legibility in the Lens, Not the Land, Do It Once, at the Right Place, Dreaming (Offline Memory Consolidation) (see-also), Overview (+2 open questions), index. Method Reflection: T5 sharpened (navigation cost as a scaling dimension) with a verdict: adopt the data model, give the agent the index, defer the database; section-level retrieval left open. Seeds evaluated → Communication and Control Are Intertwined does not bear. Evidence caveats recorded on the source page (vendor content; NICD tests graph+vector vs vector, not the wiki pattern). The article's embedded instruction to agents (install ki) was quoted, not followed.
2026-11-08 method

T5 resolved in part (human verdict): CLAUDE.md now frames the wiki as the map on top of raw/+search; Key claims must quote raw/ verbatim; "10–15 pages per source" quota dropped. New The Borges Map concept + scripts/lint/borges.py (LINT check). First run: wiki 0.70× raw (0.92× of ingested); Borges clusters: engineering craft 11.1×, Wiener 9.8×; compression proposed, tied to T1 worked examples.