Evidence Library
The public artifacts this project’s continuity claims and category map are built on. Each entry: what it is, what it supports, and what’s still open. The org-level Kenotic Labs story lives in Field Watch Entry 001 — these cards cover only the artifacts themselves. The broader field beyond this project’s own sources lives in Signals from the Field — and for the points where those sources actively conflict, where the field currently disagrees.
Lab defines what has to be tested. Evidence Library shows the receipts, references, and source material behind the claims. This supports proof-before-claim discipline; it does not mean the proof is already complete.
Evidence artifacts
ATANT v1.0 — evaluation framework paper
Continuity evaluation framework
One of the first public attempts to define and score continuity properties directly, rather than treating memory recall as a proxy for continuity.
That continuity properties (persistence, update handling, disambiguation, and similar) can be operationalized into a testable framework.
Single-author framework, single narrative corpus, self-evaluated to date on Kenotic Labs' own reference implementation (NURA) — independent replication is an open question. Published by Kenotic Labs, whose commercial work this project also profiles — read as the field's originating argument, not independent corroboration.
ATANT v1.1 — positioning paper
Continuity evaluation framework
Positions continuity evaluation explicitly against memory, long-context, and agentic-memory benchmarks, arguing continuity measures a different thing than any of them alone.
That existing memory/long-context benchmarks don't cover what continuity evaluation is trying to measure.
Single-author framework, single narrative corpus, self-evaluated to date on Kenotic Labs' own reference implementation (NURA) — independent replication is an open question. Published by Kenotic Labs, whose commercial work this project also profiles — read as the field's originating argument, not independent corroboration.
The Continuity Layer (thesis paper)
Continuity thesis / positioning
Makes the broader argument for why intelligence needs an architecture for what it carries forward, independent of any one product.
The category-level argument that carried-forward state is an architectural requirement, not an optional feature.
A positioning paper, not an empirical evaluation — read alongside the ATANT papers, which carry the testable claims. Published by Kenotic Labs, whose commercial work this project also profiles — read as the field's originating argument, not independent corroboration.
ATANT — GitHub repository
Open benchmark code
The evaluation code is public and inspectable, which is rarer than it should be in this field.
Reproducibility in principle — the code is available to run.
Public code doesn't by itself mean independent parties have run it yet; that remains open.
ATANT v1.0 corpus — HuggingFace dataset
Benchmark dataset
The narrative corpus underlying the ATANT evaluations is public and downloadable.
Transparency of the underlying test material.
Single corpus — breadth across different kinds of narratives and domains is untested.
“RAG Doesn’t Solve Hallucination. What Actually Does.”
Field commentary
Argues the specific limitation of RAG this project also maps under “Memory Is Not Continuity” — that retrieval-of-similar isn't verification-of-current.
The RAG-versus-hallucination distinction as a live, publicly argued position in the field, not just this project's framing.
Vendor-published essay — read as argument, not independent research.
“Why Do 80% of AI Agent Projects Fail?”
Field commentary
A public claim about agent-project failure rates tied to state loss between steps and sessions — relevant to the “Why This Matters Now” framing.
That agent failure is commonly attributed to lost state, not model capability, in public industry discussion.
The 80% figure is the vendor's framing; this project has not independently verified the underlying methodology.
“Why Does Character AI Forget Everything?”
Field commentary
A concrete, named-product example of the memory/continuity gap, useful as a plain-language illustration.
That consumer AI products with memory features still lose continuity across sessions in practice.
Single-product example; not a systematic survey of consumer AI memory failures.
LOCOMO
Continuity/memory benchmark
One of the earliest widely-cited benchmarks for very long-term conversational memory, testing multi-session and temporal QA.
That multi-session memory recall is measurable and that current systems have quantifiable gaps on it.
Measures conversational recall, not the fuller continuity property set (disambiguation, reconstruction, model independence) this project tracks.
LongMemEval
Continuity/memory benchmark
500-question benchmark covering information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention — the closest existing public benchmark to several of this project's seven continuity properties.
That update-handling and abstention are measurable and that current chat assistants have specific, quantified weaknesses there.
Benchmark-specific findings; results are per-model and change as models update.
BEAM
Long-context/memory benchmark
Generates coherent synthetic conversations up to 10M tokens with probing questions, designed to force genuine retrieval-based memory rather than reward long-context stuffing.
That long-context alone is distinguishable from actual memory/retrieval performance at scale.
Synthetic conversations — real-world conversational structure may differ.
MemoryBench
Memory/continual-learning benchmark
Evaluates memory and continual learning in LLM systems via simulated accumulated user feedback across multiple domains and languages.
That continual, feedback-driven memory update is measurable across domains, not just within one narrow task.
A same-named but distinct project exists at github.com/supermemoryai/memorybench — this entry refers specifically to the THUIR/ICML benchmark; do not conflate the two when citing.
RULER
Long-context benchmark
A synthetic, configurable suite measuring the real usable context size of long-context models, not just their advertised context window.
That advertised long-context length and effective usable context are frequently different things — directly relevant to the “long context isn't continuity” contrast.
Measures long-context retrieval capability, not continuity properties like update handling or model independence.
Zep evals
Vendor-published memory system evaluation
A public, technical account of a temporal knowledge-graph approach to agent memory, evaluated on DMR and LongMemEval.
That graph-structured, time-aware memory can improve measured performance on existing memory benchmarks.
Vendor-authored; a headline result reported in the paper was publicly disputed and later corrected by a third party (github.com/getzep/zep-papers/issues/5). Treat performance claims as self-reported pending independent reproduction.
Letta / MemGPT evals
Vendor-published memory system evaluation
A public leaderboard scoring models on agentic memory read/write/update tasks, from the team behind the MemGPT/Letta agent memory framework.
That agentic memory read/write/update behavior varies significantly by underlying model and is measurable.
Company-run benchmark, blog-published rather than peer-reviewed; methodology has not been independently replicated.
Once the source trail is visible, continue through evaluation standards, declared tests, local practice, or the underlying model.