The Continuity Layer
← Field Watch
Field Watch — Disagreements

Open questions in AI memory

thecontinuitylayer.com/field-watch/disagreements

Not everything in this field is settled. These comparisons connect architectural and evaluation questions to specific sources. They distinguish a difference in scope from a documented disagreement, and do not declare a winning system.

How to read this page — reviewed 6 October 2026. On that date the July version’s numerical source grades and unverified percentages were removed (see the changelog). Each entry links its sources and explains its limits. These are source-level comparisons, not independent replications.

01Emerging · 06

Emerging

Sourced comparisons of architectures or evaluation scopes. Different approaches are not necessarily opposing claims.

Benchmark validity

Emerging

Why it matters — A score is useful only when the evaluated task matches the intended use.

Approach or scope ALoCoMo and LongMemEval evaluate conversational memory through specified tasks; LongMemEval includes updates, reasoning and abstention.Approach or scope BAMA-Bench evaluates agent-environment trajectories; MemoryAgentBench evaluates memory through incremental interactions.
Primary sourceLoCoMoLongMemEvalAMA-BenchMemoryAgentBench

A broader task suite is not proof that an earlier benchmark is invalid.

Memory write strategy

Emerging

Why it matters — Compression affects storage cost, detail, later reconstruction and auditability.

Approach or scope AMem0 describes extracting, consolidating and retrieving salient information from conversations.Approach or scope BThe episodic-memory position paper emphasizes retaining instance-specific context.
Primary sourceMem0 — memory extraction and consolidationEpisodic Memory is the Missing Piece

These design emphases can coexist. Neither source establishes that all summaries destroy useful episodes or that extraction is sufficient for every task.

Retrieval mechanism

Emerging

Why it matters — Choosing what to retrieve affects which causal and temporal relationships remain available.

Approach or scope AZep describes temporal knowledge-graph memory evaluated on conversational-memory tasks.Approach or scope BAMA-Bench motivates causality-aware memory and tool-assisted retrieval for agent trajectories.
Primary sourceZep — temporal knowledge graphAMA-Bench and AMA-Agent

The sources are not a controlled head-to-head comparison and do not establish a universal winning retrieval method.

Memory representation

Emerging

Why it matters — A graph introduces relationships and maintenance costs; useful representation depends on the workload.

Approach or scope AZep develops a temporal graph representation for agent memory.Approach or scope BMem0 describes extracted memory and a graph-enhanced variant, illustrating that graph and other retrieval designs can be combined.
Primary sourceZepMem0

Comparative cost and quality require matched experiments.

Memory control plane

Emerging

Why it matters — Where memory decisions execute affects permissions, observability and failure recovery.

Approach or scope AMemGPT proposes virtual context management with movement between memory tiers.Approach or scope BCALMem describes application-layer memory and automatic context injection, alongside agent-writable facts.
Primary sourceMemGPTCALMem

This is a comparison of architectures, not evidence that all control must be inside or outside the model. CALMem also permits agent-written facts.

Forgetting policy

Emerging

Why it matters — Retention, deletion requests and retrieval quality are different requirements.

Approach or scope AFadeMem investigates selective forgetting for memory efficiency.Approach or scope BGateMem tests whether shared-memory agents honor deletion requests while preserving utility and respecting access boundaries.
Primary sourceFadeMemGateMem

Utility-driven decay is not equivalent to authorized deletion. This map does not infer a general privacy guarantee from either evaluation.

02Watchlist · 04

Watchlist

Open questions and scope comparisons. These are not established two-sided disputes.

Episodic memory architecture

Watchlist

Why it matters — What information must survive a one-off event for an agent to use it later?

Approach or scope APink and colleagues argue for an explicit, integrated episodic-memory architecture.Approach or scope BLong-context evaluation tests what remains usable inside a supplied context; it does not settle how episodes should be stored.
Primary sourceEpisodic Memory is the Missing Piece — position paperRULER — effective context evaluation

These sources address different scopes. This map has not established that RULER’s authors oppose episodic memory.

Retrieval vs long context

Watchlist

Why it matters — A design must balance context cost, useful information and access boundaries.

Approach or scope ARULER evaluates effective use of long contexts under synthetic tasks.Approach or scope BGateMem evaluates long-context and memory-based configurations under shared-memory governance tasks.
Primary sourceRULERGateMem

Neither establishes that retrieval is obsolete or always superior. Conclusions depend on workload and protocol.

Multi-agent memory

Watchlist

Why it matters — Shared state must still respect who may read, change and delete it.

Approach or scope AThread-scoped checkpoints and namespaced stores define persistence boundaries; collaboration still requires application-level policy.Approach or scope BGateMem makes shared memory across multiple principals an explicit governance evaluation problem.
Primary sourceLangGraph persistence — threads and cross-thread storesGateMem

The first statement is this map’s design question, not an attributed anti-sharing position.

Supersession vs retrieval

Watchlist

Why it matters — Returning a current answer and invalidating obsolete state are separate tests.

Approach or scope ALongMemEval includes a knowledge-update task category.Approach or scope BControl-Plane Placement Shapes Forgetting studies memory mutation operations and recovery of forgetting failures.
Primary sourceLongMemEvalControl-Plane Placement Shapes Forgetting

These are complementary evaluation scopes, not proof that LongMemEval’s authors consider retrieval alone sufficient.

03

How this list is kept

Entries are graduated onto this page from an internal 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.