AuthorsSungwoo Goo, Hwi-yeol Yun, Sangkeun Jung
TL;DR
Consolidator uses a learned slot-local phase transform at the STM–LTM boundary to route persistent memory, boosting updated-mapping recall from 18.32% to 87.02% with frozen PMNet.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Consolidator shows that without direct LTM routing, updated-mapping recall collapses to 44.38 ± 1.94%, despite 89.90% immediate STM recall.
This means PMNet with only persistent content but no routing input cannot reliably recover the updated modulo-10 mapping after a full KV and STM reset.
HOW IT WORKS
Consolidator adds a shared slot-local Consolidator, hierarchical router, routed short-term memory, and persistent long-term memory on top of PMNet’s sliding-attention KV cache.
You can think of routed STM as fast RAM, LTM as slower disk, and Consolidator as a tiny controller that decides how to rewrite disk at each boundary.
This KEY_MECHANISM lets Consolidator preserve and revise address-specific rules across context resets, something a plain Transformer context window cannot do.
DIAGRAM
This diagram shows how Consolidator processes two context segments, consolidates routed STM into LTM twice, clears KV and STM, and then answers the final query from persistent LTM.
DIAGRAM
This diagram shows how Consolidator compares learned versus identity consolidation and routing on versus routing off while freezing different parameter subsets.
PROCESS
Hierarchical routed writes
Consolidator uses the hierarchical router to write phase-valued state into routed short-term memory slots during each segment, while KV cache handles local attention.
Shared slot-local phase transform
At the boundary, Consolidator applies the shared slot-local phase transform Cψ to each occupied STM slot using cosine and sine coordinates.
Persistent accumulation and reset
Consolidator accumulates transformed STM into long-term memory, then clears the KV cache, STM, and occupancy masks while retaining LTM across segments.
Direct LTM-conditioned routing
In later segments and during the final query, Consolidator feeds LTM into same-level routing state so the hierarchical router selects slots based on persistent experience.
KEY CONTRIBUTIONS
Shared slot-local transform for routed STM
Consolidator introduces a shared slot-local phase transform with 12.35K parameters that consolidates routed short-term memory into long-term memory without replaying source tokens.
Direct LTM-conditioned slot routing
Consolidator injects persistent long-term memory directly into the hierarchical router, raising updated-mapping recall from 44.38% to 87.02% (+42.64 pp) while STM recall stays 89.90%.
Sequential same-address update task
Consolidator is evaluated on a two-segment modulo-10 mapping task that separates carrying state, revising at a reused address, retrieving LTM, and using LTM to guide slot selection.
RESULTS
Updated mapping LTM recall
87.02%
+68.70 pp over forced identity in Consolidator-only
Updated mapping LTM recall
44.38%
-42.64 pp vs Consolidator-only with direct routing
Segment 2 STM recall
89.90%
matched across routing on and off conditions
Trainable parameter fraction
0.041%
12.35K of 29.95M parameters trained in Consolidator-only
On the two-segment modulo-10 same-address update benchmark, Consolidator demonstrates that a tiny boundary operator can enable persistent, revisable memory in PMNet. The MAIN_RESULT shows that learned consolidation plus direct LTM routing recovers updated mappings far better than identity accumulation or routing-off baselines.
BENCHMARK
On the two-segment modulo-10 same-address update benchmark, Consolidator demonstrates that a tiny boundary operator can enable persistent, revisable memory in PMNet. The MAIN_RESULT shows that learned consolidation plus direct LTM routing recovers updated mappings far better than identity accumulation or routing-off baselines.
BENCHMARK
Updated-mapping LTM recall after second consolidation for learned versus identity consolidation in the Consolidator-only setting.
BENCHMARK
Updated-mapping LTM recall with learned versus identity consolidation, comparing routing on and routing off conditions.
KEY INSIGHT
Consolidator boosts updated-mapping LTM recall from 44.38% without direct LTM routing to 87.02% with routing, while STM recall stays fixed at 89.90%.
This is surprising because many would expect persistent content alone to suffice, but Consolidator shows that feeding LTM into routing decisions adds a 42.64 percentage point gain.
WHY IT MATTERS
Consolidator shows that a tiny, shared boundary operator can turn routed STM into a persistent access state that shapes future slot selection.
With Consolidator, builders can bolt persistent, revisable memory onto frozen architectures like PMNet, enabling forward-only adaptation across context resets without retraining or replay.
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.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…