AuthorsMinkyu Song
TL;DR
Dependency-aware Semantic Garbage Collection (DSGC) propagates one-hop prerequisite support, boosting target full-chain retention from 0.03 to 0.90 and 0.23 to 1.00.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Structurally indirect prerequisite eviction causes full-chain retention to collapse to 0.03 under a lexical encoder when similarity-only policies face budget pressure.
In long-horizon agentic memory tasks, this eviction breaks the reasoning chain before retrieval, so downstream retrievers and LLM agents cannot recover necessary evidence.
HOW IT WORKS
Dependency-aware Semantic Garbage Collection (DSGC) combines semantic relevance ri, a one-hop propagation term πi, a directed prerequisite graph E, and greedy construction of the retained subset S under a fixed budget.
Think of DSGC like a tracing garbage collector for agent memory: query-facing blocks are roots, and liveness propagates one hop along prerequisite edges instead of raw pointer links.
This one-hop propagation lets DSGC keep structurally necessary but weakly query-aligned blocks live, preserving reasoning chains that a plain context window with similarity-only ranking would evict.
DIAGRAM
This diagram shows how Dependency-aware Semantic Garbage Collection separates retention-stage failures from downstream retrieval failures in agentic memory.
DIAGRAM
This diagram shows how the paper constructs synthetic scenarios, applies retention policies, and measures full-chain retention for Dependency-aware Semantic Garbage Collection.
PROCESS
Compute softmax relevance ri
Dependency-aware Semantic Garbage Collection embeds the query and each block, then computes normalized ri with a fixed temperature τ = 0.25 across the memory store.
Propagate one hop via prerequisite graph E
Using the directed prerequisite graph E, DSGC sums downstream ri scores into πi for each immediate prerequisite, defining structural support for vulnerable blocks.
Combine scores into Ii and rank blocks
Dependency-aware Semantic Garbage Collection adds ri and λπ πi to form Ii, then sorts blocks by this combined score while preserving deterministic seeded order on ties.
Greedy selection of retained subset S under budget C
DSGC greedily builds the retained subset S by adding blocks in descending Ii until the token budget C is exhausted, defining the prompt-visible store for retrieval.
KEY CONTRIBUTIONS
Operational definition of structurally indirect prerequisite eviction
Dependency-aware Semantic Garbage Collection formalizes structurally indirect prerequisites via simϕ(q, bi) < simϕ(q, bj) and distinguishes retention-stage failure from retrieval and reasoning failures.
Deterministic benchmark and trace diagnostics
Dependency-aware Semantic Garbage Collection is evaluated on a fixed 20-block synthetic benchmark with 4 templates, 15 seeds, and per-seed traces decomposing displaced blocks, competitors, and score margins.
One-hop dependency-aware retention rule
Dependency-aware Semantic Garbage Collection introduces a one-hop propagation rule Ii = ri + λπ πi that raises full-chain retention from 0.03 to 0.90 and 0.23 to 1.00 in target regimes.
RESULTS
Full-chain retention lexical target
0.90
+0.87 over similarity-only lexical baseline
Full-chain retention sentence target
1.00
+0.77 over similarity-only sentence baseline
Full-chain retention lexical control
1.00
matches similarity-only and no-graph DSGC ceilings
Full-chain retention sentence control
1.00
shows no degradation from propagation
The fixed main suite benchmark uses 4 templates, 3 policies, and 2 encoders with 15 seeds per condition to test structurally indirect prerequisites under budget pressure.
These results show that Dependency-aware Semantic Garbage Collection almost completely repairs retention-stage failures in target regimes while preserving perfect retention on retrieval-friendly controls.
BENCHMARK
The fixed main suite benchmark uses 4 templates, 3 policies, and 2 encoders with 15 seeds per condition to test structurally indirect prerequisites under budget pressure. These results show that Dependency-aware Semantic Garbage Collection almost completely repairs retention-stage failures in target regimes while preserving perfect retention on retrieval-friendly controls.
BENCHMARK
Full-chain retention over 15 seeds for lexical and sentence encoders on target scenarios.
KEY INSIGHT
Even with a strong sentence encoder, similarity-only retention still drops full-chain retention to 0.23, while Dependency-aware Semantic Garbage Collection reaches 1.00.
This is surprising because dense embeddings are often assumed to fix retrieval, yet DSGC shows that structural prerequisite propagation is still necessary before retrieval begins.
WHY IT MATTERS
Dependency-aware Semantic Garbage Collection gives agent builders a concrete way to treat retention as a structural reachability problem, not just a similarity ranking problem.
With DSGC, developers can design memory systems that keep entire reasoning chains live under budget pressure, enabling more reliable long-horizon agents and mechanistic failure analysis before retrieval.
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…