AuthorsSaber Zerhoudi, Jelena Mitrovic, Michael Granitzer
TL;DR
Knowledge Triage uses typed operators TypeCompact, TypeDecompose, and TypeRetrieve to eliminate the Compaction Cliff, preserving up to 1.00 constraint recall vs 0.53 for Sonnet 4.6 /compact at 50% compression.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Under Claude Code’s /compact prompt on Sonnet 4.6, safety-rule recall is 53% after one compaction round and only 10% after five.
This failure in long-running AI agents lets critical constraints like drug allergies vanish, causing unsafe recommendations and broken safety guarantees over extended sessions.
HOW IT WORKS
Knowledge Triage centers on a five-type classifier SafetyMargin, plus three deterministic operators TypeCompact, TypeDecompose, and TypeRetrieve over a typed knowledge base.
Like triage in a hospital separating critical from stable patients, Knowledge Triage separates constraints, procedures, and softer items before deciding how to compress or move them.
This KEY_MECHANISM of per-type distortion lets Knowledge Triage preserve safety rules exactly while still shrinking or partitioning other memory, something a plain context window or uniform compactor cannot do.
DIAGRAM
This diagram shows how Knowledge Triage composes TypeCompact, TypeDecompose, and TypeRetrieve into a single context-management loop.
DIAGRAM
This diagram shows how Knowledge Triage is evaluated on compaction, decomposition, retrieval, and behavioral benchmarks using six corpora.
PROCESS
Typed Knowledge Model Construction
Knowledge Triage builds the tuple K = (I, τ, T, π, σ), assigning each item a type and scope so TypeCompact, TypeDecompose, and TypeRetrieve can reason over constraints.
Per-item Type Classification
SafetyMargin and the selective cascade classify items into Constraint, Procedural, Belief, Preference, and Episodic, providing the labels that drive Knowledge Triage’s per-type retention.
TypeCompact With Guard Verifier
TypeCompact routes constraints and procedures into a hard lane, compresses softer items, then runs the guard verifier to ensure every constraint survives or flags Unsafe.
TypeDecompose and TypeRetrieve
When budgets are tight, TypeDecompose replicates scoped constraints into partitions, and TypeRetrieve pins in-scope constraints ahead of relevance for queries during the session.
KEY CONTRIBUTIONS
The Compaction Cliff
Knowledge Triage identifies the Compaction Cliff where Sonnet 4.6 /compact drops safety-rule recall from 53% after one round to 10% after five on 20 agent configurations.
Five-type Knowledge Model and SafetyMargin
Knowledge Triage defines five operational types covering 97% of AgentArtifactCorpus and introduces SafetyMargin, lifting declarative constraint recall from 0.62 to 0.90 across phrasing forms.
Typed Operators With Formal Guarantees
Knowledge Triage’s TypeCompact, TypeDecompose, and TypeRetrieve achieve 2–4× higher constraint recall than single-shot compactors, 0% locality violations vs 93%, and 100% recall@50 vs 73% for Sonnet 4.6.
RESULTS
Constraint preservation at 50%
1.00 recall
+0.47 over Sonnet 4.6 (/compact) which has 0.53 recall
Constraint preservation at 25%
0.95 recall
+0.56 over LLMLingua-2 which has 0.39 recall
Locality violations in decomposition
0% violations
-93 percentage points vs chunk_by_paragraph which has 93% configs with violations
Retrieval recall@50
100% recall@50
+27 percentage points over Sonnet 4.6 one-shot which has 73% recall@50
On AgentArtifactCorpus compaction experiments, Knowledge Triage’s TypeCompact preserves all classifier-flagged constraints at 50% compression, while Sonnet 4.6 /compact preserves only 0.53.
On BEIR scifact plus τ-bench retail policies, TypeRetrieve reaches 100% in-scope constraint recall@50, proving that Knowledge Triage can maintain safety rules under aggressive retrieval budgets.
BENCHMARK
On AgentArtifactCorpus compaction experiments, Knowledge Triage’s TypeCompact preserves all classifier-flagged constraints at 50% compression, while Sonnet 4.6 /compact preserves only 0.53. On BEIR scifact plus τ-bench retail policies, TypeRetrieve reaches 100% in-scope constraint recall@50, proving that Knowledge Triage can maintain safety rules under aggressive retrieval budgets.
BENCHMARK
Constraint recall at 50% compression across prompt compactors on 20 AAC configurations.
BENCHMARK
Pass rate on τ-bench retail tasks comparing full policy, hierarchical truncation, and TypeCompact.
KEY INSIGHT
Despite compressing aggressively, Knowledge Triage’s TypeCompact maintains 1.00 constraint recall at 50% compression, while Sonnet 4.6 /compact collapses to 0.10 after five rounds.
This is surprising because production prompts explicitly say “keep every safety rule,” yet type-blind summarization still deletes rules, showing that natural-language instructions alone cannot enforce safety preservation.
WHY IT MATTERS
Knowledge Triage unlocks long-running agents whose safety rules survive repeated compaction, decomposition, and retrieval without being paraphrased away.
Builders can now deploy agents with bounded context that still obey critical constraints over hours of interaction, using typed memory operators instead of hoping the context window and prompts are enough.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…