Research digest · COLM 2026 · Context Beyond the Window

The Recall Debt

Flat Memory Schemas Structurally Fail Multi-Hop Retrieval

One result · visually

Evidence before interpretation.

Similarity retrieval collapses before bridge-aware retrieval does.

BM250%

Dense0%

Stronger similarity≤22.6%

Bridge resolved61.6%

The same object · seven altitudes

From question
to falsifier.

A paper becomes easier to inspect when its claim, boundary, intervention, measurement, failure mode, and next disconfirming result are kept separate.

01Question
Can a flat memory retrieve evidence whose relevance only appears through an intermediate concept?
02Claim
Some apparent forgetting is structural: the retrieval schema contains no traversable path to the needed evidence.
03System boundary
Document retrieval and candidate selection—not the downstream model’s ability to answer from supplied evidence.
04Intervention
Resolve the intermediate bridge before asking the retriever to surface the terminal document.
05Measurement
Recall@10 on 159 gold-annotated two-hop MuSiQue chains, plus answer accuracy when the missing terminal is supplied.
06Failure mode
Embedding-disjoint candidates can look like bridges without being semantically valid; disjointness is not a ground-truth label.
07Next falsifier
A flat retriever that consistently recovers the residual bridge cases without an explicit traversal structure.

The problem in plain language

Two conversations can be about the same underlying project without sharing the words that make that relationship obvious. A dense retriever cannot rank what is not nearby in embedding space; a lexical retriever cannot match vocabulary that never overlaps.

Once both routes fail, adding a larger embedding model or tuning a top-k value does not repair the missing path. The system needs another way to move through memory.

What the experiments show

On 159 gold-annotated two-hop MuSiQue chains in the embedding-disjoint tail, BM25 and dense retrieval recover the terminal document 0% of the time at recall@10. When the correct terminal is supplied, a long-context model answers 47.5% correctly versus 5.0% from standard retrieval.

This is not only a weak-retriever artifact: stronger open and frontier embedding systems remain at or below 22.6%, while resolving the single intermediate bridge reaches 61.6%.

Flat memory makes some multi-hop relationships unreachable by construction.
Candidate selection can fail even when the downstream model can answer from the missing evidence.
The schema needs a path to the intermediate bridge; a stronger similarity score is not enough.

Why I care about it

An agent that remembers individual facts but cannot reconnect discoveries across threads accumulates recall debt. The archive grows while the ability to compound knowledge does not.

This is also a provenance problem: a recovered claim should carry the path by which it became relevant, not arrive as an isolated fragment with no lineage. The operational-corpus audit also exposes a measurement trap: only 2% of the mined disjoint-tail candidates were judged genuine bridges, so disjointness alone is not a validity gate.

Still open

The remaining question is whether other structures—graph retrieval, temporal links, or learned traversal policies—recover the residual cases without erasing the evidence path that made them trustworthy.