Dated, source-backed profiles from the continuity field map. Entries are reference points, not rankings or endorsements.
Dated profiles of the labs, tools, benchmarks and systems this map re-checks. For newer or lighter-touch items, see Signals.
Continuity research lab
Publishes the ATANT continuity evaluation framework and a thesis about carrying state across AI hosts.
The vendor describes a local engine with deterministic extraction and cross-host MCP support. These are vendor claims; this project has not tested the implementation.
Field commentary
Samuel Sameer Tanguturi, founder of Kenotic Labs, argues that similarity retrieval alone does not establish current state. This is an author-attributed architectural argument, not independent verification.
Essay by Samuel Sameer Tanguturi, published by Kenotic Labs on 7 April 2026. Kenotic works in a field where this site’s maintainer also does commercial work; the projects are unaffiliated. Read as the author’s argument, not independently validated research.
Continuity/memory benchmark
Evaluates long-term conversational memory through question answering, event summarization and multimodal dialogue generation.
Its tasks extend beyond fact recall. A score on this suite does not by itself establish permission enforcement, recovery or transfer between independent systems.
Continuity/memory benchmark
The authors publish 500 questions across information extraction, multi-session reasoning, temporal reasoning, knowledge updates and abstention.
These skills overlap with continuity concerns, but results depend on model, dataset variant and evaluation protocol. This project has not reproduced reported scores.
Long-context benchmark
A synthetic, configurable suite measuring the real usable context size of long-context models, not just their advertised context window.
A synthetic suite covering retrieval, multi-hop tracing and aggregation. It tests effective context use, not a complete governance or cross-system recovery protocol.
Frontier model lab — context & memory features
Provides models, context-management features and a memory tool whose storage is implemented by the client application. The storage and its permissions belong to the surrounding system, not the model weights. Since 23 April 2026, Claude Managed Agents also offer hosted memory stores in public beta: memories are stored as files that can be exported, shared across agents with read-only or read-write scopes, and tracked in an audit log. In the Claude app, memory from chat history has been available on all plans, including free, since 2 March 2026.
Documented memory and context primitives, now including hosted memory stores for Managed Agents (public beta) and app memory on every plan. Application design still determines persistence, access control and recovery; this project has not tested the hosted stores.
Frontier model lab — memory & stateful APIs
Provides stateful conversation APIs for applications. ChatGPT memory is a separate product feature; neither should be described as persistent memory in the underlying model weights. OpenAI's release notes add memory sources on all consumer plans (5 May 2026), which show saved memories and past chats that may have shaped a response, and a memory summary page (in place by June 2026); since 12 June 2026 its Delete and turn off memory control removes what it lists. OpenAI notes that sources may not show every factor behind a response.
Official documentation listed 26 August 2026 as the Assistants API shutdown date; that date has passed. This source review does not independently test endpoint availability. Consult the Responses and conversation-state documentation for migration.
Agent framework — memory & persistence modules
LangGraph checkpointers persist thread state; stores support memory across threads. Persistence also supports interrupts, resumption and fault recovery.
Framework capabilities extend beyond state within a single workflow. Permissions, provenance and project-level authority still depend on the application.
Enterprise context infrastructure (temporal context graphs)
Zep now describes itself as “the unified context layer for enterprise data”: its Context Lake unifies business data, documents and conversations into context graphs for agents. Its use of “context layer” overlaps with this site's vocabulary; the term is Zep's own and implies no affiliation. Zep's Community Edition is deprecated; the open-source Graphiti framework (Apache-2.0) is the self-hosted option. Its January 2025 paper remains a public, technical account of a temporal knowledge-graph approach to agent memory, evaluated on DMR and LongMemEval.
The January 2025 paper evaluates DMR and LongMemEval and reports no LoCoMo result. Separately, a Mem0 cofounder disputed Zep's blog LoCoMo figure of 84% and claimed 58.44% (issue #5, May 2025). Zep acknowledged a calculation error and published 75.14%, while rejecting Mem0's figure. All three numbers are vendor-reported. These are reported LoCoMo question-answer accuracy figures from differing evaluation setups, not a common-model leaderboard. A matched model/configuration for all three figures was not independently established here. The dispute is evidence status, not an assessment that the product is inactive.
Agent harness with persistent memory (Letta Code)
Letta's 16 March 2026 post “Our next phase” makes Letta Code, “a model-agnostic agent harness with persistent memory”, its flagship. Memory moves from database-editing tools into git-backed files (context repositories, also called MemFS), and Letta Filesystem was set to be deprecated by mid-April 2026. History: Letta's May 2025 leaderboard article describes agentic memory read, write and update tasks, and its separate August 2025 filesystem evaluation reports 74.0% LoCoMo accuracy with GPT-4o mini and constrained tool use.
Letta's current focus is Letta Code with git-backed memory files, per its March 2026 post; this project has not tested it. The May and August 2025 evaluations are historical, company-run and not a current ranking; they describe different evaluations, and this project has not reproduced their methodology or results.
AI platform and memory models
The vendor now describes an AI acceleration platform with SDKs, memory models and long-horizon workflows, making it relevant beyond a shared name.
Product and benchmark claims are vendor-published. Model versions, dataset variants and protocols must be matched before comparing scores; this project has not validated the service.
AI project-continuity system
Astrolobe describes a product for carrying technical-project state, unresolved failures and validation evidence across sessions, tools, models and handoffs.
The vendor's store page is live and describes Continuity 1.0.1. Product capabilities are vendor claims; this project has not tested the package.
Memory exchange and project handoffs
The vendor describes PMX (Protocol for Memory Exchange), project state, tasks, handoffs, provenance and explicit approval, alongside personal memory.
Its stated scope includes project continuity. Implementation and enforcement claims have not been independently tested by this project.
Open-weight model family — memory lives in your application
Meta publishes Llama weights that anyone can download and run under the Llama 4 Community License. Llama 4 Scout and Llama 4 Maverick were released on 5 April 2025; the model card lists a context length of 10 million tokens for Scout and 1 million for Maverick, and describes each as a static model trained on an offline dataset. The weights carry no memory between sessions: whatever an assistant built on Llama remembers is stored and governed by the application around it. Meta's open-source agent server, Llama Stack, is now named OGX; its repository describes an MIT-licensed, OpenAI-compatible API server with tool calling, MCP server integration and file search. It does not describe a memory feature.
Open weights under Meta's own community license. Memory, if any, is built by whoever runs the model. This review covers the Llama 4 model card and the OGX repository; it has not tested either, and later Meta releases are not reviewed here.
Open-weight model family — agent toolkit, memory in your application
Alibaba's Qwen team states that all its open-weight Qwen3 models are licensed under Apache 2.0. The Qwen3 series was released on 29 April 2025, and the July and August 2025 updates (Qwen3-2507) support 256K tokens of context, extendable to 1 million. The team describes the models as built for agent work with external tools. Its Qwen-Agent toolkit (Apache 2.0) adds function calling, MCP tool configuration and retrieval over long documents. As with any open weights, memory across sessions is something the application stores, not the model.
Open weights under Apache 2.0, with an open agent toolkit beside them. This review covers the Qwen3 and Qwen-Agent repositories; it has not tested them, and newer Qwen releases are not reviewed here.
Open-weight model family — stateless API
DeepSeek publishes model weights under the MIT License; its DeepSeek-R1 repository lists a 128K context length for DeepSeek-R1. DeepSeek's API guide says the chat API is stateless: the server does not record the context of a user's requests, so the application must send the whole conversation history with every request. That makes the memory question plain: with DeepSeek, continuity is whatever your own application keeps and sends.
Open weights under the MIT License and a stateless chat API: the application carries all history. This review covers the DeepSeek-R1 repository and the API guide; it has not tested them, and newer DeepSeek models are not reviewed here.
Open-weight model family — stateful Agents API
Mistral AI publishes downloadable models alongside its hosted ones; its mistral-inference repository (now archived) lists models such as Mistral 7B, Mixtral 8x7B and 8x22B, Mistral Nemo and Mistral Small 3.1, and states that all the listed models support function calling. Licenses vary by model: the repository names custom non-commercial licenses for Codestral 22B and Mistral Large 2. On the hosted side, Mistral's Agents API, announced on 27 May 2025, keeps stateful conversations whose context and history persist across interactions, with built-in connectors for code execution, web search, image generation and MCP tools. That hosted history lives with Mistral, not in the open weights.
Open weights for some models, custom licenses for others, and a hosted Agents API that keeps conversation history. This review covers the two sources listed; it has not tested them, and newer Mistral releases are not reviewed here.
Papers, datasets, repos and articles the map tracks with a status line and sources, but no profile page of their own. Each has a stable anchor on this page.
Continuity evaluation framework
Kenotic Labs defines continuity properties and proposes a framework for evaluating them directly, rather than using recall alone as a proxy.
Author-reported continuity framework and evaluation claims. This project has not independently reproduced the results; a source review is not replication.
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.
Author-reported continuity framework and evaluation claims. This project has not independently reproduced the results; a source review is not replication.
Continuity thesis / positioning
Makes the broader argument for why intelligence needs an architecture for what it carries forward, independent of any one product.
Kenotic Labs' positioning paper — distinct from this site, despite the shared name — not an empirical evaluation; read alongside the ATANT papers, which carry the testable claims.
Evaluation specification and artifacts
Publishes the evaluation standard, examples, result files and a LoCoMo analysis script, making parts of the reported work inspectable.
The v1.2 README names a runner in a separate Nura reference implementation repository; a complete public runner was not verified in this review. The ATANT repository license reserves rights and permits reading, reference and evaluation; the Hugging Face corpus separately uses CC-BY-4.0. Public artifacts are not a complete reproduction package.
Reported evaluation artifacts
The authors report Nura passing 4,160 of 5,000 questions (83.2%) across 500 stories using the v1.0 substring-matching grader. Failed answers are separately classified by keyword cosine similarity against a distractor baseline; that classification is not the pass-rate metric.
The v1.2 README describes this 500-story, 5,000-question run. The root README separately reports 1,761 of 1,835 questions (96%) in a 250-story cumulative run. Corpus size and question count differ; scores are not interchangeable. The v1.2 paper remains listed as arXiv:TBD. Neither run was reproduced here.
Benchmark dataset
The narrative corpus underlying the ATANT evaluations is public and downloadable.
Transparent test material; a single corpus — breadth across narrative kinds and domains is untested.
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.
The 80% figure is the vendor's framing; this project has not independently verified the underlying methodology.
Field commentary
A concrete, named-product example of the memory/continuity gap, useful as a plain-language illustration.
Single-product example; not a systematic survey of consumer AI memory failures.
Long-context/memory benchmark
Tests memory over synthetic conversations up to 10 million tokens, with 2,000 validated questions across ten abilities in the authors’ release.
The suite includes reasoning and temporal or preference-related abilities as well as retrieval. Synthetic-data results and context-length settings require qualification; this project has not reproduced them.
Memory/continual-learning benchmark
Evaluates memory and continual learning in LLM systems via simulated accumulated user feedback across multiple domains and languages.
THUIR announced an llm-as-judge branch on 7 September 2026 with unified scoring. Record the branch and evaluator when comparing results. This is distinct from supermemoryai/memorybench.
Local-first AI workspace and memory tools
The current README describes an alpha desktop workspace with chat, projects, notes, journals, plugins and an MCP memory server; continuity-memory is one part of that broader project.
Project continuity is within its stated scope. The alpha label and public code describe development status, not independently verified handoff reliability.
Research architecture and prototype
The repository proposes a three-tier memory architecture and an Echo proof of concept. It also advances a consciousness hypothesis, which this map does not endorse.
The README describes the full system as under development. Treat it as a research/prototype artifact, not evidence of a completed operational system.
Independent commentary
Antoine's May 2026 essay argues that rapid model launches and deprecations disrupt users' accumulated context, working habits and long-running projects.
The essay was readable in the browser on 6 October 2026. It is commentary, not an empirical evaluation; its numerical and industry-wide claims are not independently verified here. Archive availability has not been established.
Forgetting and deletion evaluation
Dongxu Yang's June 2026 preprint studies the operations that change stored memory (supersede, release and purge) rather than recall, across thirteen system configurations. It introduces ForgetEval, including intent-aware deletion cases, and argues that existing benchmarks measure only recall.
Single-author preprint with author-reported results. This project has not reproduced ForgetEval or the reported deletion accuracy.
Context governance and verifiable erasure
Context Layer Systems, Inc., founded by Morgan Allen, describes CCL as “a post-authentication context governance and infrastructure layer” and an open architectural standard, with STACCR as its pre-GA implementation. The site says only encrypted token references pass through the system and that destroying a subject's key is designed to make that context inaccessible, backed by a signed erasure certificate.
The site describes STACCR as pre-GA and says CCL is not an AI memory store. Its name is close to this site's; the projects are unaffiliated, and this entry makes no claim about who first used the name. Capability claims are the vendor's; this project has not tested them.
Shared-memory governance evaluation
The June 2026 preprint evaluates shared memory across multiple principals, combining legitimate-request utility, access boundaries and forgetting after deletion requests.
Source scope reviewed; benchmark results have not been reproduced by this project. Evaluation of shared-memory behavior is not a certification of an entire institutional governance system.
Where the sources in this field compare, differ or leave questions open.
Dated comparisons of architectures and evaluation scopes, each with its primary sources and a note on what it does not establish. No winning system is declared.