The Guarded Memory Plane: Agentic Memory for Long-Horizon Workflows

Agent platforms routinely combine three different capabilities: durable execution state, retrieval over a document corpus, and memory. On a LangGraph-based platform, the checkpointer provides only the first. It saves a full state snapshot at each step and supports thread isolation, resume, replay, and fork. That makes it working memory with durability, scoped to a single thread. The cross-thread store beside it has no provenance, no validity window, no conflict resolution, and no handling rules for personal data. This report describes what to build on top. It separates three meanings of “memory tier” that the literature uses interchangeably: content type, access locality, and representation format. Treating them as one idea leads to poor architecture, because only the last one decides whether a stored fact can still be inspected, redacted, or deleted. The report then specifies the memory lifecycle as a guarded plane. Typed stores sit behind a write guard, and every retrieval passes through a read guard. Every record carries provenance and bi-temporal validity. Retention is bounded and scored by utility, supersession is deterministic, and retrieval is hybrid and iterative. On tiering, the report follows the published evidence. A study of twelve memory systems across five workloads finds no dominant architecture. A search policy over raw, uncompressed history also outperforms compression-based tiering on the benchmark that tiered systems cite as their own evidence. The recommendation is to start flat and add tiering only as a measured, A/B-tested decision with an explicit trigger condition. The report applies the same approach to four failure modes: context rot, lossy-summarization drift, contextual-integrity violations that grow with usage, and memory poisoning. Poisoning is a separate risk class, because a written record becomes a privileged input on the next turn and input guardrails never see it. The report closes with a build sequence ordered by retrofit cost and a measurement plan.

Authors

Publication Details

Journal
Preprints.org
Published
2026-11-05
DOI
https://doi.org/10.20944/preprints202611.0572.v1
Primary Topic
Advanced Database Systems and Queries
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

The Guarded Memory Plane: Agentic Memory for Long-Horizon Workflows

Ashwini Kumar Sharma
Preprints.org
Advanced Database Systems and Queries
preprint

The Guarded Memory Plane: Agentic Memory for Long-Horizon Workflows

Ashwini Kumar Sharma
preprint en

Abstract

Agent platforms routinely combine three different capabilities: durable execution state, retrieval over a document corpus, and memory. On a LangGraph-based platform, the checkpointer provides only the first. It saves a full state snapshot at each step and supports thread isolation, resume, replay, and fork. That makes it working memory with durability, scoped to a single thread. The cross-thread store beside it has no provenance, no validity window, no conflict resolution, and no handling rules for personal data. This report describes what to build on top. It separates three meanings of “memory tier” that the literature uses interchangeably: content type, access locality, and representation format. Treating them as one idea leads to poor architecture, because only the last one decides whether a stored fact can still be inspected, redacted, or deleted. The report then specifies the memory lifecycle as a guarded plane. Typed stores sit behind a write guard, and every retrieval passes through a read guard. Every record carries provenance and bi-temporal validity. Retention is bounded and scored by utility, supersession is deterministic, and retrieval is hybrid and iterative. On tiering, the report follows the published evidence. A study of twelve memory systems across five workloads finds no dominant architecture. A search policy over raw, uncompressed history also outperforms compression-based tiering on the benchmark that tiered systems cite as their own evidence. The recommendation is to start flat and add tiering only as a measured, A/B-tested decision with an explicit trigger condition. The report applies the same approach to four failure modes: context rot, lossy-summarization drift, contextual-integrity violations that grow with usage, and memory poisoning. Poisoning is a separate risk class, because a written record becomes a privileged input on the next turn and input guardrails never see it. The report closes with a build sequence ordered by retrofit cost and a measurement plan.

Preprints.org
Advanced Database Systems and Queries
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon