p-research

Frontier Deepresearch: In-Model Memory

2026-08-19 · PRDT Deepresearch Series №2 · generated from the P-Research corpus pipeline Method: corpus retrieval + 90-day targeted arXiv sweep + GraphRAG communities + top-venue odds. All numbers are re-derivable from data/papers.jsonl; abstract-level evidence is a research signal, not a verified fact.

Dedicated to Attention Is All You Need (Vaswani et al., 2017) — the starting point of the memory research traced in this note.

Lineage: from Attention Is All You Need to today

2017   Attention Is All You Need     attention = the algebraic form of memory
  ↓    the transformer era           parameters as memory
  ↓    KV-cache engineering          memory's ledger (cost and restoration become the theme)
  ↓    long context                  the capacity race for memory
  ↓    2025–2026 neural memory modules    memory as an explicit architectural component (MoNe / memory trees / memory forests)
  ↓    agent memory                  memory's consumer (cross-session, restorable, auditable)

Key numbers

Metric	Value
Memory-related papers (of the 574-paper corpus)	95
Full-text verified	8
Year distribution	2023: 20 → 2024: 22 → 2025: 17 → 2026 (half year): 36

Observation 0 — Memory research sharply accelerated in 2026. A single half year produced 36 papers, more than any previous full year in the corpus.

Findings

1. Memory × RAG fusion is the dominant theme (co-occurrence 51)

Memory/RAG is the #1 co-occurring method family among Memory papers in the corpus. Read together with the RAG report of this series (#1), the picture is convergent: retrieval memory and in-model memory are growing into the same thing. On the public side, the two discussion zones that map to them (#3 RAG, #4 Memory) will become progressively harder to separate — a reflection of the boundary dissolving, not a taxonomy failure. (Method-tag co-occurrence is an abstract-level signal.)

2. Modular neural memory — a new architectural category

• MoNe: Modular Neural Memory for Efficient Long Context Inference — modular neural memory for efficient long-context inference.
• ArborMem: Navigating Interaction States with Memory Forests — navigating interaction states with “memory forests”.
• Dynamic Compression in Recurrent Networks — dynamic compression of recurrent state.

Trend (abstract-level): memory is moving from a byproduct of parameters to an explicit, pluggable architectural component.

3. Restoration and persistence become themes

Bounded-State Restoration: Decoupling Local Restore Capacity from External LLM State decouples local restore capacity from KV-cache state. Memory state management — save / restore / cross-session — is becoming a first-class citizen, which directly echoes the persistence requirements of agent harnesses, including the P-Research program’s practice around agent-harness state persistence.

4. Memory × reasoning depth

• Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling — memory lets reasoning depth escape rote memorization.
• BATON: Long-Horizon Robot Manipulation via Agentic Subtask Exploration and Transition-aware Memory — transition-aware memory for long-horizon robot tasks.

Two of the top three Memory papers by top-venue odds are memory-enabled-reasoning work — the strongest crossover signal in this sweep.

5. Memory × multimodality

VLM co-occurrence: 19. Memory Tree Guided Key Frame Querying (3D QA) and MemCatalyst (VLM data auditing) both put memory structures — trees, key frames, audits — to work beyond plain text. Abstract-level signal: memory is moving past text-only settings.

6. Efficiency is destiny (as the graph structure confirms)

In GraphRAG community 0 (the 115-paper efficiency community), Memory/RAG co-resides with KV Cache, MoE, Quantization, and Speculative Decoding. Memory research always carries a cost constraint — that is what distinguishes it from other capabilities.

Top-venue odds (Memory papers)

BATON (0.609) · HAF (0.555) · Beyond Memorization (0.543) · Neuro-Symbolic Local RetoMaton (0.532) — memory-enabled reasoning/embodied papers are sitting in the top-venue range. Odds are model estimates over corpus records; treat them as research signals, not predictions.

Convergence with Report #1 (RAG)

The RAG report found retrieval memory climbing; this Memory report finds in-model memory accelerating. Together: the two legs of agent memory are both sprinting, and they are fusing into one.

Implications for the P-Research program

Element	Frontier signal
Memory	Modular neural memory + state restoration are the two hotspots
RAG	Fusing with Memory (co-occurrence 51) — the boundary between the two zones is dissolving
Planning	Memory forests (ArborMem) = planning over memory structure
Eval	MemCatalyst uses memory for data auditing — memory can be an evaluation tool
Tool Use	Weak signal this round; deferred to a dedicated sweep

Program-position statements:

• The program’s practice in agent-harness persistence (save / restore / cross-session state) is aligned with the restoration thread above: the external literature and the program’s practice are converging on the same requirement — memory state management as a first-class component of agent systems.
• The five capability zones that form the public curriculum of PAT’s five elements (RAG / Memory / Tool Use / Planning / Eval) map 1:1 onto the frontier signals in this sweep; the Memory and RAG zones in particular should be read together rather than apart.
• The method behind this note — corpus retrieval, a 90-day targeted arXiv sweep, GraphRAG communities, and top-venue odds, with full-text verification of 8 Memory papers — is what surfaced these signals, and keeps every abstract-level claim auditable against data/papers.jsonl.

Open problems

1. Where are the evaluation benchmarks for neural memory modules (the MoNe class)?
2. Can memory state restoration (the Bounded-State Restoration class) become a standard component of agent harnesses?
3. Financial time-series memory — what shape does “market memory” take? This is the Quant × Memory crossover, and the ground where the program’s Quant × AI pillar sits.

One-line takeaway

Retrieval memory and in-model memory are both accelerating — and converging into a single agent-memory substrate.

This site is open source. Improve this page.