---
title: More documents, worse answers: retrieval gets its mechanics audited
url: https://hangingcontext.com/stories/2026-10-15/rag-mechanics/
published: 2026-10-15T12:46:07Z
edition: 2026-10-15T12
newsroom: artificial-intelligence
beat: research
audience: ai_developer
trust: reported
source_count: 4
cited_manifest_ids: ['1792051595354213082', '1792051595356359184', '1792051595356604559', '1792051595356646833']
---

# More documents, worse answers: retrieval gets its mechanics audited

*Four new preprints take RAG's machinery apart: adding documents degrades retrieval through vector search dilution, frozen external memory needs a target-aligned reader, a dual-process memory cuts output tokens 73.91%, and consensus-driven evidence selection trims runtime 75.9%.*

Reported · 4 sources · Artificial Intelligence · Published 2026-10-15 12:46 UTC

## Key facts

- In a preprint, the authors report that vector search dilution causes significant performance degradation in RAG systems as document collections scale, that hard partitioning of document indices is unsafe for RAG retrieval, and that training-free soft scoping is a safe and effective method for domain-scoped retrieval. (self-reported) [cites: 1792051595354213082]
- In a preprint, the authors report that frozen memory artifacts are not inherently model-agnostic and require a target-aligned reader to be effective, and that a dual-layer, four-branch reader architecture effectively bridges the performance gap between same-model and cross-model memory reuse. (self-reported) [cites: 1792051595356359184]
- In a preprint, MemDuet's authors report an end-to-end inference speedup of 9.74x and an output-token reduction of 73.91% relative to RL-MemAgent, using a dual-process memory framework with System 1 for selection and System 2 for verification. (self-reported) [cites: 1792051595356604559]
- In a preprint, ConRAG's authors report that consensus-driven multi-view ranking improves multi-hop QA evidence selection, and that ConRAG reduces end-to-end runtime by 75.9% and token consumption by 67.3% on Gemma-4-31B-it. (self-reported) [cites: 1792051595356646833]

## Bigger corpus, worse retrieval

The headline finding of the batch: vector search dilution causes significant performance degradation in RAG systems as document collections scale. Add documents, lose precision and recall — the opposite of the intuition that more knowledge helps. The authors' prescription: training-free soft scoping is a safe and effective method for domain-scoped retrieval, while hard partitioning of document indices is unsafe for RAG retrieval. A second finding from the same paper: multi-agent orchestration in RAG systems triggers a precision-faithfulness paradox that is resistant to standard prompt and format controls. [cites: 1792051595354213082]

The paradox finding is the sleeper — our read. Orchestrating multiple agents makes answers more precise but less faithful, and no prompt tweak fixes it. That is a warning for the agent-orchestration enthusiasm: adding agents can make the groundedness problem worse, not better. Design the system for the trade-off; do not expect a prompt to erase it. [cites: 1792051595354213082]

## Frozen memory needs the right reader

A second preprint asks whether external memory can be reused across models — and answers: only with the right reader. Frozen memory artifacts are not inherently model-agnostic and require a target-aligned reader to be effective. The authors' fix is a dual-layer, four-branch reader architecture that effectively bridges the performance gap between same-model and cross-model memory reuse. And frozen memory can provide utility without target-side training if the reader interface is compatible — knowledge storage decoupled from model weights, with a compatibility tax. [cites: 1792051595356359184]

> **From Synorb** — Ask what Synorb covers. The MCP tool synorb-stream-search answers questions like "do you have this publisher?" without using any Manifest allowance. [Connect and ask](https://synorb.com/connect)


## Two systems are better than one

The third preprint, MemDuet, borrows from human cognition: a dual-process memory framework with System 1 for selection and System 2 for verification, which the authors say improves the accuracy-efficiency trade-off in long-context reasoning. The numbers: MemDuet achieves an end-to-end inference speedup of 9.74x and an output-token reduction of 73.91% relative to RL-MemAgent. Training MemDuet using GRPO for memory rewriting and verification improves long-context reasoning performance, the authors report. [cites: 1792051595356604559]

## Consensus beats more passes

The fourth preprint, ConRAG, targets multi-hop QA. It improves multi-hop QA performance by using consensus-driven multi-view ranking for evidence selection, and its slot-based mechanism effectively transfers intermediate information between dependent reasoning steps. The efficiency numbers: ConRAG reduces end-to-end runtime by 75.9% and token consumption by 67.3% on Gemma-4-31B-it. [cites: 1792051595356646833]

## Our read

Our read: retrieval is becoming an engineering discipline with its own failure taxonomy — dilution, reader mismatch, selection-verification trade-offs, evidence fragmentation. The practical throughline: the fixes are all about scope and structure (soft scoping, aligned readers, slot-based state, consensus ranking), not bigger models. [cites: 1792051595354213082, 1792051595356359184, 1792051595356646833]

The most actionable of the four for practitioners is probably the dilution finding — our read. If your RAG got worse as the corpus grew, that is now a named phenomenon with a recommended fix: soft scoping, not hard partitioning. The most ambitious is cross-model memory reuse — a world where knowledge artifacts outlive the model that wrote them. The paradox finding, meanwhile, is a design constraint for the multi-agent RAG boom: orchestrating more agents buys precision at the cost of faithfulness, and the paper says standard prompt and format controls do not fix it. Plan the architecture around it, not the prompt. The honest caveat: four preprints, author-reported numbers, and the dilution paper's real-world evidence is not quantified in the extracted claims. [cites: 1792051595354213082, 1792051595356359184]

## What is still unresolved

- At what corpus size vector search dilution becomes operationally significant — the claims report the phenomenon, not a threshold.
- Whether the dual-layer, four-branch reader generalizes beyond the evaluated model pairs.
- All four are preprints with no independent replication yet.

## Decision card

- **Useful for:** RAG engineers debugging retrieval degradation at scale; Teams designing cross-model memory or multi-hop QA systems
- **Not for:** Anyone looking for a drop-in fix — these are research findings, not shipped features
- **Compare against:** Hard-partitioned index baselines; BF16-rollout and RL-MemAgent baselines reported in each paper

## Provenance ledger

| Manifest id | Title | Source | Claims used | Published |
|---|---|---|---|---|
| `1792051595354213082` | When More Documents Hurt RAG: Measuring and Reducing Vector Search Dilution with Domain-Scoped Retrieval | [arxiv-rag-search-knowledge](https://arxiv.org/pdf/2606.11350v2) | 1792052038954725333, 1792052038957478964, 1792052038963218084, 1792052038965913740 | 10-15 00:00 |
| `1792051595356359184` | Frozen Memory Is Not Enough: Rethinking External Memory as Extraction | [arxiv-rag-search-knowledge](https://arxiv.org/pdf/2608.17050v4) | 1792052037574159749, 1792052037592025929, 1792052037645631260 | 10-15 00:00 |
| `1792051595356604559` | MemDuet: Learning Dual-Process Memory for Long-Context Reasoning | [arxiv-rag-search-knowledge](https://arxiv.org/pdf/2610.19928v1) | 1792052050779304627, 1792052050784058715, 1792052050786196900 | 10-15 00:00 |
| `1792051595356646833` | ConRAG: Consensus-Driven Evidence Flow for Multi-Hop Question Answering | [arxiv-rag-search-knowledge](https://arxiv.org/pdf/2605.28093v3) | 1792052040680565222, 1792052040684313668, 1792052040682748848 | 10-15 00:00 |

## Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

1. **Get a Synorb Key.** Open https://synorb.com/connect in a browser, or let your agent provision one: MCP tool `synorb-get-demo-key` (no header needed), or REST `GET https://synorb.com/connect` with `Accept: application/json`. Save the key. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; 3 of them are starter Beacons.
2. **Read each cited Manifest by id.** REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
3. **Check the claim.** Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.

Cited Manifest ids:

- `1792051595354213082`
- `1792051595356359184`
- `1792051595356604559`
- `1792051595356646833`

```sh
for id in 1792051595354213082 1792051595356359184 1792051595356604559 1792051595356646833; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done
```

> **From Synorb** — Run the cited Manifests yourself. Every claim in this paper points to a Manifest id. Get a Synorb Key and read the Manifests behind it. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; 3 of them are starter Beacons. [Get a key](https://synorb.com/connect)

## Related stories

- [The KV cache goes on a diet: INT2 quantization nears lossless](https://hangingcontext.com/stories/2026-10-15/kv-cache-diet/)
- [Inference is being re-priced and re-placed, one token at a time](https://hangingcontext.com/stories/2026-10-15/inference-new-map/)
- [Web agents now face time travel, unified shells, and distracting facts](https://hangingcontext.com/stories/2026-10-15/web-agent-evals/)
- [Your coding agent can ship features. It probably can't keep them secure.](https://hangingcontext.com/stories/2026-10-15/vibe-code-security/)

## Links

- HTML: https://hangingcontext.com/stories/2026-10-15/rag-mechanics/
- JSON: https://hangingcontext.com/stories/2026-10-15/rag-mechanics/index.json
- Markdown: https://hangingcontext.com/stories/2026-10-15/rag-mechanics/index.md

Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.

---
Reuse (terms v1): Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product. Credit: "Hanging Context, with provenance by Synorb", with a link to https://hangingcontext.com/stories/2026-10-15/rag-mechanics/. Terms: https://hangingcontext.com/terms/reuse/