AuthorsEvan Chen, Shiqiang Wang, Christopher G. Brinton
TL;DR
PLANFENCE uses dependency-scoped lineage validation at the action boundary to prevent stale-plan execution, keeping invalid actions at 0/330 while freshness-only baselines issue 330/330 invalid actions.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Distributed LLM-agent teams can read the latest shared facts yet still act on an obsolete plan, causing 330/330 invalid actions under freshness-only policies.
In reservation, fulfillment, and deployment workflows, an executor may hold a revised requirement but still call tools using arguments from a stale plan, leading to cancelled orders or obsolete deployments.
HOW IT WORKS
PLANFENCE combines Exact derivation, Declared scope, Authoritative currency, and a Dependency-scoped action gate so plans cite precise public record IDs and executors validate only action-relevant keys.
You can think of PLANFENCE like a transaction commit check for tools: instead of refreshing the whole database, it validates just the specific records that authorize a pending external call.
This lineage-focused gate lets PLANFENCE guarantee that a protected action descends from current public evidence, something a plain context window or freshness-only memory cannot establish.
DIAGRAM
This diagram shows how PLANFENCE validates declared dependencies with owners immediately before executing a protected action, including replan and block paths.
DIAGRAM
This diagram shows how PLANFENCE and baseline memory policies are evaluated across controlled workflows, update schedules, and network traces.
PROCESS
Recording action-relevant lineage
PLANFENCE attaches exact parent IDs to each derived public record, so requirements and role decisions record the precise versions that produced them.
Declaring dependency scope
Tool wrappers in PLANFENCE declare D(a), the semantic keys whose values can affect the protected action, forming a trusted dependency contract.
Dependency-scoped action gate
At action time, PLANFENCE traverses exact parents, queries owners for H(x), compares the dependency frontier Fa with heads, and decides authorize, replan, or block.
Plan repair and revalidation
If PLANFENCE detects a mismatch, it fetches updated records, triggers one replan, and then revalidates; any second change or incomplete response causes the action to fail closed.
KEY CONTRIBUTIONS
Identifying stale-plan execution and lineage validity
PLANFENCE formalizes stale-plan execution and lineage validity, showing that freshness-only checks issue 330/330 invalid actions while exact-lineage policies issue 0/330 invalid actions.
Design of dependency-scoped PLANFENCE gate
PLANFENCE introduces a dependency-scoped action gate that binds plans to exact public inputs, validates only tool-declared dependencies, and permits one replan before failing closed.
Coordination stall boundary across policies
PLANFENCE maps when proactive synchronization or action-time validation is cheaper, showing lower stall and traffic than batched all-key validation across 8–128 key settings.
RESULTS
Invalid actions per issued
0/330 actions
-330/330 versus Local replica
Availability per scheduled
330/330 actions
matches Centralized lineage availability
Coordination stall
230.8 ms/action
177.8 ms lower than Centralized lineage on AT&T
Distributed traffic
8.1 KiB/action
7.5 KiB lower than Batched all-key validation at 8 keys
These numbers come from the eight-key compact-state workload on the pinned AT&T trace at high churn, with 42 updates and 11 protected actions per episode. They show that PLANFENCE can maintain zero invalid actions while reducing coordination stall and traffic compared with freshness-only baselines and safe all-key validation policies.
BENCHMARK
These numbers come from the eight-key compact-state workload on the pinned AT&T trace at high churn, with 42 updates and 11 protected actions per episode. They show that PLANFENCE can maintain zero invalid actions while reducing coordination stall and traffic compared with freshness-only baselines and safe all-key validation policies.
BENCHMARK
Invalid actions and coordination stall per action for PLANFENCE and baseline memory policies.
KEY INSIGHT
Even with perfectly fresh replicated state, Local replica and Owner-head freshness each issue 330/330 actions from obsolete plans, while PLANFENCE issues 0/330 invalid actions.
This is surprising because many systems assume that fresh memory is enough; PLANFENCE shows that without explicit derivation lineage, executors can still act on stale authorization despite current state.
WHY IT MATTERS
PLANFENCE unlocks lineage-safe tool execution in distributed LLM-agent teams, ensuring that protected actions are authorized by current, action-relevant public records.
Builders can now design agent workflows that tolerate high churn and large shared keyspaces without global synchronization, validating only the dependencies that matter for each external call.
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…