The Continuity Layer
Field Watch — Disagreements

Where the field currently disagrees

thecontinuitylayer.com/field-watch/disagreements

Not everything in this field is settled. This page tracks the live disagreements — where credible sources currently pull in different directions on how AI memory and continuity should work. It resolves none of them. Each entry shows both positions, their primary sources, and how well-evidenced each side is: a record of where the field stands, not a verdict on who is right.

How to read this page — two independent signals sit on every entry. The disagreement status (contested, emerging, or watchlist) is how live and two-sided the split is. The source grade (★) rates the quality of the evidence behind each side — it is not a claim that the side is correct. A better-sourced position is better-evidenced, not proven.

01Contested · 07

Contested

Live, two-sided splits with credible evidence on both sides. The field is actively arguing these.

Episodic memory architecture

Contested

Why it matters — Determines whether agent/project memory needs a distinct episodic store or rides existing retrieval.

ClaimEpisodic memory (single-shot, instance-specific, context-bound) is THE missing piece for long-term LLM agents — an explicit integrated focus is needed now.Extending context length + existing RAG/vector retrieval is the primary path; no distinct episodic-memory architecture is required.
Primary sourcearXiv:2502.06975 (Pink et al., Position: Episodic Memory is the Missing Piece, Max Planck MPI-SWS)long-context extension literature (state-space / context-window-extension camp)
Source grade4 of 5source grade3 of 5source grade

Better-sourced side: Position A — a source-grade statement (better-evidenced, not a verdict on which side is correct).

Benchmark validity

Contested

Why it matters — If the standard benchmarks don't measure real memory use, teams optimize for the wrong target.

ClaimStandard memory benchmarks (LoCoMo, LongMemEval) measure long-term memory adequately via conversational recall.Passive-recall benchmarks mislead — models that ace LoCoMo drop to 40-60% on task-embedded memory (MemoryArena); recall != decision-relevant memory use.
Primary sourceLoCoMo (Maharana et al. 2024), LongMemEval (Wu et al. 2025) — de facto standard evalsvalidity critique: MemoryArena (He et al. 2026) via survey arXiv:2603.07670 ('passive recall aces are poor memory agents'); incremental-eval critique arXiv:2507.05257 (LoCoMo too short, LongMemEval synthetic/low-diversity)
Source grade3 of 5source grade4 of 5source grade

Better-sourced side: Position B — a source-grade statement (better-evidenced, not a verdict on which side is correct).

Memory write strategy

Contested

Why it matters — Summarize-on-write trades recall fidelity and auditability for efficiency — a core continuity-layer design fork.

ClaimSummarizing memories at write time is efficient and sufficient for agent memory.Summarizing at write time collapses distinct episodes into semantic generalizations, destroying the episodic signal (Tulving binding) before it can be used — raw episodic detail must be preserved.
Primary sourcecommon summary-on-write agent-memory practice (Mem0 arXiv:2504.19413 and similar)episodic-preservation argument (Pink et al. 2502.06975; Tulving binding principle)
Source grade3 of 5source grade3 of 5source grade

Both sides are comparably sourced.

Retrieval mechanism

Contested

Why it matters — Whether retrieval must model causal/temporal structure or similarity suffices changes the whole retrieval stack.

ClaimSimilarity-based and entity-centric retrieval (vector search, GraphRAG entity graphs) is adequate for agent memory retrieval.Existing methods neglect the underlying causality within stored information; retrieval must model causal/temporal structure, not just similarity.
Primary sourceGraphRAG (Edge et al. 2025) / HippoRAG2 (Gutierrez et al. 2025) entity-centric retrievalarXiv:2602.22769 (AMA-Bench, causality critique §2.2); graph-retrieval evidence: HopRAG (Liu et al. 2025, ACL Findings, +36% multi-hop acc), GNN-RAG (Mavromatis & Karypis 2024)
Source grade3 of 5source grade3 of 5source grade

Both sides are comparably sourced.

Memory representation

Contested

Why it matters — Whether to build and maintain a graph layer or ship vector search — a major cost and complexity decision.

ClaimStructured knowledge-graph memory is needed for multi-hop relational reasoning; flat vector search misses relationships.Graph memory adds construction cost, complexity, and latency most agent use cases don't need; vector retrieval is the pragmatic default.
Primary sourcegraph camp: HopRAG (Liu et al. 2025 ACL, +36% multi-hop), GNN-RAG (Mavromatis & Karypis 2024, +8.9-15.5% F1), Microsoft GraphRAG (Edge et al. 2025)vector-default camp: 2026 framework comparisons (vectorize.io, Atlan) — adds complexity/latency, graph features often paywalled/overkill
Source grade3 of 5source grade3 of 5source grade

Both sides are comparably sourced.

Memory control plane

Contested

Why it matters — Decides where memory logic lives — the model's own tool calls or your app — shaping the whole system architecture.

ClaimThe LLM should manage its own memory — self-editing, OS-style paging between context and external store.Memory management belongs outside the model in application logic; delegating it to the LLM is burdensome, adds latency, and hasn't proven out in production.
Primary sourceMemGPT (Packer et al. 2023, arXiv:2310.08560) / Lettaapplication-layer camp: CALMem (arXiv:2605.20724); practitioner critique (Towards Data Science agent-memory guide, Apr 2026 — burdensome, tends to fail, no production use)
Source grade3 of 5source grade3 of 5source grade

Both sides are comparably sourced.

Forgetting policy

Contested

Why it matters — Whether your memory system prunes or accumulates — directly affects long-run retrieval quality, cost, and privacy.

ClaimAgent memory should selectively forget — retention-only strategies poison retrieval precision over time; forgetting is essential for robustness, privacy, efficiency.Most systems retain everything within capacity (binary retain-or-lose); selective forgetting is largely unsolved and risks deleting safety-critical information.
Primary sourceforgetting-as-feature: memory-frontiers survey arXiv:2603.07670 ('forgetting is not a bug, it is a feature'); FadeMem arXiv:2601.18642 (biologically-inspired decay, 45% storage cut)retain-default + risk: same survey ('most systems fail conspicuously on selective forgetting'; 'learned forgetting could delete safety-critical information')
Source grade3 of 5source grade3 of 5source grade

Both sides are comparably sourced.

02Emerging · 02

Emerging

Newer fault lines. One side is often the field's working default rather than a defended, published position.

Retrieval vs long context

Emerging

Why it matters — Determines whether you invest in a retrieval/memory layer at all, or bet on context windows growing.

ClaimLong context windows are making external retrieval/memory systems unnecessary — just put everything in the prompt.RAG remains necessary and superior for production: bounded cost, filtered attention, knowledge bases exceeding any context window.
Primary sourcelong-context camp (multiple 2025-26 surveys and aggregations)RAG-production camp (context-engineering literature; AMA-Bench arXiv:2602.22769 §2.2)
Source grade2 of 5source grade3 of 5source grade

Better-sourced side: Position B — a source-grade statement (better-evidenced, not a verdict on which side is correct).

Multi-agent memory

Emerging

Why it matters — Whether agents each keep private memory or converge on shared state — the core multi-agent continuity architecture question.

ClaimEach agent maintains its own memory/continuity (per-agent state).Multi-agent systems need shared or converged memory; per-agent private memory causes contamination, divergence, and coordination failure.
Primary sourceper-agent-memory default (single-agent frameworks extended to teams)shared/converged camp: GAM (Wu et al. 2026, write-isolation for contamination) via arXiv:2605.27864; GateMem multi-principal shared-memory arXiv:2606.18829
Source grade2 of 5source grade3 of 5source grade

The map holds this one neutral — both positions are tracked, and it favors neither.

03Watchlist · 01

Watchlist

Early or one-sided splits we are tracking but not yet calling contested.

Supersession vs retrieval

Watchlist

Why it matters — If stale facts are never actually removed, memory silently accumulates contradictions over time.

ClaimKeeping the latest fact retrievable is what 'knowledge update' means — surfacing the current value is sufficient handling of changed facts.Retrieving the latest fact is not enough — the superseded fact must actually be removed/invalidated, or stale facts silently accumulate; systems fail this (<=28% on multi-hop fact-consolidation).
Primary sourcelatest-fact-retrieval framing (LongMemEval knowledge-update category design)supersession critique: FORGETEVAL arXiv:2606.15903 ('asks whether latest fact is retrieved, not whether superseded fact removed'); FactConsolidation results arXiv:2507.05257 (<=28% multi-hop)
Source grade2 of 5source grade3 of 5source grade

Better-sourced side: Position B — a source-grade statement (better-evidenced, not a verdict on which side is correct).

04

How this list is kept

Entries are graduated onto this page from an internal, continuously updated review of the field, and this page renders only that published record — so what you read here is deploy-time stable and citable. How entries are chosen, categorized, and re-checked is set out in How the map works. Source quality follows the project’s editorial policy.