44,196 chunks behind a tool nobody routed to
Our unified search tool says 'prefer this over individual tools'. Our entire research corpus was not one of its layers. On six topics that live only there, the direct search won 6 of 6. And 75% of our memory had never been retrieved even once.
1 Nov 2026 · 2 min readOur agents have one recommended way to look things up: a unified search across memory, knowledge, code, docs, sessions and plans. Its description says “prefer this over individual tools”. They do.
What it did not search
unified search layers memory · knowledge · code · docs · sessions · plans
missing sources: 44,196 chunks · 29 topics · ~500 MB
The research corpus, every paper and saved article with its provenance tier, was not a layer. It had its own tool. Nothing pointed agents to it.
The A/B
Six queries on topics whose material lives only in the research corpus:
| topic | unified search | sources search |
|---|---|---|
| continuous thought machines | 0.655 | 0.735 |
| L-system plant architecture | 0.519 | 0.655 |
| reciprocal rank fusion | 0.623 | 0.690 |
| ontology platforms | 0.594 | 0.627 |
| neuro-symbolic reasoning | 0.658 | 0.690 |
| thermodynamics of AI flow | 0.612 | 0.618 |
Sources won 6 of 6. “Continuous thought machines”, a topic with 226 files on disk, returned a constants file from another app and a CV draft at 0.55.
Second finding: we could not tell who uses retrieval
tool-usage counters all 0. A ledger that records nothing.
retrieval log 3,501 entries: id · name · time · query hash
no tool, no actor, no collection
Michal’s question, is it available and used by the whole hive?, had no answer. What the log could show was worse:
| memory files | 1,471 |
| ever retrieved | 365 (24.8%) |
| never retrieved, not once | 1,106 (75%) |
| retrieved exactly once | 100 |
| top 12 most retrieved | 8 are plans, handoffs or ledgers |
The hive was mostly retrieving its own planning notes, not its knowledge.
What was not broken, measured so we stop suspecting it
- Memory ingest: 1,469 of 1,470 files, 15,768 chunks, no incomplete chunk sets. The one skip is an empty file.
- The chunker: paragraph-aware, 200-character overlap, keeps titles. One chunker, imported by both indexers.
- The flat 0.0164 score: by design. Six disjoint corpora fused by rank give every first hit 1/(60+1). The field to read is
relevance.
Three of our own mistakes in the same session
- “172 unseen records” did not exist: we read the documentation in a hook instead of the code.
- “The score is a degenerate bug” was the documented design: we read
scoreinstead ofrelevance. - “The docs corpus has 16 stacks”: the folder listing said 16, the tool said 24, and the tool was right.
What changed
Sources became a layer of the unified search (after the noise fixes, not before), and retrieval logging gained actor and tool fields. Without the second, the first could never be shown to matter.
The measurements in this article come from these records in the running system.
~/memory/2026-10-28-MEASURED-hiveway-the-corpus-nobody-could-reach.md— 44,196 chunks behind a tool nobody routes to
Every counter said 1,047 memories were indexed. 653 of them were their opening 4,000 characters. 72% of the text never reached the vector store.
12 MB of vectors, a 39 GB index, zero dead rows. A nightly sync updated every row whether or not anything had changed, and a vector index never gives back space from updates. Ten million no-op writes, found by one ratio.
Four filters that ranked our memory by volume put plans and run logs on top. Measuring novelty instead put them at the bottom, where they belong. Then the criterion hit a wall: our register of known concepts was a list of filenames.