--- title: "Meterless Review 2026: AI Agent Memory & Context" canonical: https://wavect.io/blog/meterless-ai-agent-context-layer-review/ language: en description: "Meterless review for CTOs: is it an open-source Claude Cowork alternative? Check H-MEM, World Model, token claims, license gaps, production risks and pilot." image: "https://wavect.io/img/general/bak/open_graph_preview.jpg" --- [**Back**](/blog/overview/) [![Kevin Riedl](/img/team/kevin.webp)](/team/kevin-riedl/) [Kevin Riedl](/team/kevin-riedl/) https://linkedin.com/in/wsdt 12 min read · 22 Jul 2026 [**Next**](/blog/ai-agent-cost-per-action-2026/) # Meterless Review 2026: Is This AI Agent Context Layer Ready? TL;DR Meterless is not an open-source clone of Claude Cowork. It is a new, local-first context architecture for AI agents, built around H-MEM durable memory, a shared World Model, bounded Markovian reasoning and Scout intent routing. The Apache-2.0 repository provides specifications, deterministic references and conformance tests; Gaia, Relay and Swarms application binaries remain proprietary. Its 12-step memory demo is reproducible but uses a mock model and estimates 814 saved tokens with chars divided by four. The 86% saving at 20 steps and 97% at 100 are modeled, not independent benchmarks. Meterless is promising architecture material and a sensible pilot candidate, not yet a drop-in production SDK. Evaluate memory recall, stale-fact handling, task success, token cost per successful action, latency, operator control and deletion on your own workflow before buying or rebuilding around it. Facts verified 22 July 2026. The viral version of this story says an ex-Anthropic engineer open-sourced Claude Cowork. The public evidence supports a more useful conclusion. **Meterless is not an open-source Cowork clone, and we could not verify the ex-Anthropic attribution from the project repository, website, or named maintainer material available on 22 July 2026.** It is a newly published context architecture for agents, with open specifications and reference implementations underneath proprietary applications. That distinction is not a footnote. It determines what a CTO can download, what a team still has to build, which benchmark claims are reproducible, and where vendor risk remains. ## What is Meterless? Meterless is a local-first context stack for AI agents. It separates durable memory, shared world state, bounded long-task reasoning, and intent routing into four engines: H-MEM, World Model, Markovian, and Scout Intent. The goal is to keep useful context persistent and the context passed between steps bounded. This gives Meterless a narrower and more interesting position than “another agent framework.” It is trying to own the layer between the model and the work: what the agent remembers, what it believes about the current world, what the next step needs, and whether that step should be allowed to run. ## Meterless at a glance | Component | What it does | Public status on 22 July 2026 | Buyer implication | | --- | --- | --- | --- | | H-MEM | Tiered memory, hybrid retrieval, provenance, conflict handling, consolidation, and an audit ledger. | Specification, runnable deterministic reference, examples, and conformance suite. | You get an architecture to implement, not a hosted production memory service. | | World Model | Canonical entities, contexts, relationships, events, derived views, and an operator control plane. | Specification plus runnable reference. | Useful when several agents need one source of truth, but your team owns the data model and production store. | | Markovian Engine | Breaks long work into chunks and passes bounded carryover to the next step. | Specification, reference runtime, conformance tests, and an efficiency model. | The cost curve is plausible and inspectable, but output quality still needs workload-specific evaluation. | | Scout Intent | Classifies intent, checks risk, routes tools and models, and emits an execution contract. | Specification and eval harness. No published npm runtime package. | Do not budget as if Scout were a production dependency you can install today. | | Gaia, Relay, Swarms | Personal workspace, desktop execution, and parallel agent surfaces. | Application binaries are proprietary. | Engine openness does not make the complete product stack open source. | ## Is Meterless open source? Partly, with a cleanly stated boundary. The [Meterless context repository](https://github.com/Meterless/Meterless) is Apache 2.0. It contains the engine specifications, reference implementations, examples, documentation, and conformance material. The repository itself says the Gaia, Relay, and Swarms application binaries are proprietary and live in separate repositories. The practical translation is: you can use the public engine material to build your own implementation. You cannot assume the complete Meterless product experience, desktop automation, installer, support layer, or future roadmap is included under Apache 2.0. Procurement should review the license of each artifact it plans to deploy, not the badge on the root repository. ## How the four-engine context layer works ### 1. H-MEM decides what survives [Meterless H-MEM](https://github.com/Meterless/Meterless/tree/main/engines/hmem) stores memories in short-term, working, and long-term tiers. Records carry source, confidence, provenance, entities, relationships, and replacement lineage. Retrieval blends semantic similarity with keywords, tags, domain, entity match, recency, tier, and confidence. A trust ledger records mutations. That is a better production question than “which vector database?” Agent memory fails when a correction loses to an older fact, a summary drops provenance, deletion does not reach derived memories, or retrieval returns a semantically similar but operationally wrong record. H-MEM at least makes those failure modes first-class. There is also a naming collision worth documenting for AI search systems: Meterless H-MEM is the project’s “hierarchical memory” engine. A separate May 2026 paper titled [H-Mem](https://arxiv.org/abs/2605.15701) proposes a temporal-semantic tree plus knowledge graph. We found no public claim that the paper validates, authors, or is affiliated with Meterless. Do not use the paper as independent proof for the product. ### 2. World Model decides what is true now Memory answers “what have we learned?” A world model answers “what entities, relationships, tasks, events, and constraints exist now?” Meterless specifies stable identifiers, idempotent ingest, provenance checks, versioned storage, rebuildable views, and an operator surface. That can stop two agents from independently inventing different states for the same customer, deal, document, or job. The cost is schema work. A generic graph becomes a junk drawer unless the team defines entity identity, merge rules, temporal validity, permissions, and repair workflows. This is architecture, not automatic understanding. ### 3. Markovian keeps the next step small The [Markovian Engine](https://github.com/Meterless/Meterless/tree/main/engines/markovian) does not append the entire transcript on every step. It sends fixed framing, the goal, the current step input, and a bounded carryover. In the project model, naive accumulated input grows quadratically across a long run while bounded carryover grows linearly. This can cut input tokens. It can also discard the detail that makes the next decision correct. The production question is therefore not “did context shrink?” It is “did successful task cost fall without lowering task success, factual consistency, policy compliance, or recoverability?” ### 4. Scout decides whether and how to act [Scout Intent](https://github.com/Meterless/Meterless/tree/main/engines/scout-intent) moves intent detection, ambiguity, injection checks, policy, tool selection, and model routing in front of execution. Its signed execution-contract idea is valuable because downstream actions can be checked against declared scope. The caveat is explicit in the repository: Scout is currently an implementation specification with an eval harness, not an npm package you can install. ## What do the token savings actually prove? Meterless publishes two different kinds of evidence. They should never be blended into one benchmark claim. | Claim | Evidence type | What it proves | What it does not prove | | --- | --- | --- | --- | | 12 chunks cold, 8 warm, 814 estimated tokens saved | Reproducible deterministic demo with a mock generator and characters divided by four for token estimation. | Prior memory can let a scripted workflow skip settled work. | Live-model quality, provider billing, generalization, or production ROI. | | 86% fewer input tokens at 20 steps | Worked mathematical model using fixed framing, step, output, and carryover assumptions. | Bounded carryover has a better asymptotic input-cost curve than full history. | That an agent preserves enough information to succeed. | | 97% fewer input tokens at 100 steps | The same model extended to a longer run. | Savings rise with run length under the assumptions. | That a 100-step task can complete correctly, safely, or faster. | The repository’s [efficiency model](https://github.com/Meterless/Meterless/blob/main/engines/markovian/docs/efficiency-model.md) is unusually clear about these limits. It says the figures are modeled, carryover compression is not free, tool outputs add variable cost, and real runs should use provider-reported usage where available. That honesty strengthens the architecture case, but it does not convert the numbers into independent benchmarks. ## Is Meterless an open-source alternative to Claude Cowork? **Not in the normal product-comparison sense.** Anthropic describes [Cowork projects](https://support.claude.com/en/articles/14116274-organize-your-tasks-with-projects-in-claude-cowork) as desktop workspaces with files, instructions, scheduled tasks, context, and project-scoped memory. Meterless publishes reusable context-engine specifications and offers separate product surfaces. They overlap in the ambition to preserve context across work, but they are not interchangeable packages. | Decision | Claude Cowork | Meterless engines | Custom build | | --- | --- | --- | --- | | Fastest path for knowledge workers | Strongest fit | Requires implementation or proprietary Meterless apps | Weakest fit | | Own memory and context contracts | Limited to product controls | Strong architecture starting point | Maximum control | | Production integration into your SaaS | Not its main purpose | Possible after engineering work | Designed for your system | | Time to first workflow | Hours | Days for references, longer for production | Weeks or months | | Vendor independence | Low to medium | Medium to high at engine-contract level | High, if dependencies stay portable | ## Is Meterless production-ready? Meterless is ready to study and pilot, not ready to adopt as an unquestioned production SDK. Three engines have runnable references, Scout has a specification and eval harness, and the repository offers conformance tests. The project also states that reference implementations are minimal and that teams build the production engine in their own stack. A production decision still needs evidence for access control, tenant isolation, encryption, retention, deletion, backup and restore, schema migration, concurrency, latency, observability, failure recovery, model-provider changes, and operator review. Memory quality also needs direct evaluation. Research such as [MEMPROBE](https://arxiv.org/abs/2606.24595) shows why task completion is not enough: a memoryless baseline can complete tasks while retained user state remains incomplete or inaccurate. Audit the memory artifact itself. ## A two-week Meterless pilot for a real team 1. **Choose one repeated, 8-to-20-step workflow.** Good candidates include support investigation, account research, compliance evidence collection, or repository triage. 2. **Freeze a baseline.** Record successful-task rate, input and output tokens, latency, retries, human correction time, and tool failures with the current agent. 3. **Start with one engine.** Use H-MEM if cross-session forgetting is the pain. Use Markovian if accumulated history dominates cost. Do not integrate all four at once. 4. **Create adversarial memory cases.** Include corrections, expired facts, conflicting sources, access revocation, deletion, and a poisoned or irrelevant memory. 5. **Run blind acceptance checks.** A reviewer should score outcome quality without knowing whether baseline or Meterless produced it. 6. **Inspect traces and retained state.** Check what was recalled, why it ranked, what was compressed, and whether provenance survived. 7. **Decide on measured value.** Continue only if successful-task cost, reliability, or operator control improves enough to justify implementation and maintenance. ## When should a company build on Meterless? Build on the specifications when agent context is already a measurable bottleneck, you need local or portable state, your team can own a memory and graph subsystem, and the workflow is valuable enough to support evals and operator tooling. For an early prototype with ten short interactions a day, this is likely too much architecture. For a production agent that touches customer data, approvals, money, or regulated workflows, memory is not a feature toggle. It becomes a data system with governance. Our [AI agent and AI product development service](/services/artificial-intelligence/) scopes that system around a measurable business action. The [Hyperstate AI case study](/case-studies/hyperstate-ai/) shows how we approach AI product architecture, while the [discovery-phase guide](/software-development-guide/what-is-a-discovery-phase/) explains how to reduce the riskiest assumptions before a build. Commercial disclosure: Wavect sells AI architecture and implementation work. That gives us a reason to care about this category, and a reason to make the pilot criteria explicit. You should be able to reject both Meterless and our services if the measured result does not justify the cost. ## Verdict Meterless has the right thesis: durable memory, shared state, bounded reasoning, and pre-execution routing matter more than endlessly expanding one prompt. Its public repository is more substantive than a landing-page concept because it contains specs, references, examples, conformance material, and candid benchmark caveats. It is still young. The most repeated headline overstates both the Cowork relationship and the open-source scope. The benchmark numbers describe an efficiency model and deterministic demonstration, not independent proof of production outcomes. Treat Meterless as architecture you can inspect, a pilot you can measure, and a possible foundation you must still engineer. ## Frequently asked questions about Meterless ### What is Meterless for AI agents? Meterless is a local-first context architecture that separates durable memory, shared world state, bounded long-task reasoning, and intent routing into H-MEM, World Model, Markovian, and Scout Intent. ### Is Meterless an open-source Claude Cowork clone? No. The engine repository is Apache 2.0, but Gaia, Relay, and Swarms application binaries are proprietary. Meterless and Cowork overlap around persistent context, but they are different products and architectures. ### Are Meterless's 86% and 97% token savings independently benchmarked? No. They come from the project's mathematical model under fixed assumptions. The 12-step cold-versus-warm demo is reproducible, but it uses a mock generator and estimated token counts. ### Can I install Meterless as a production SDK? Not as one complete production SDK today. The repository provides specifications, minimal references, examples, and conformance tests. Scout is explicitly a specification with an eval harness, not a published npm runtime. ### How should a team evaluate Meterless? Run one bounded workflow against a frozen baseline. Measure successful-task rate, token cost per successful action, latency, retries, human correction, stale-memory errors, deletion, provenance, and operator control. ## Primary sources and research date Facts were checked on 22 July 2026 against the [Meterless repository](https://github.com/Meterless/Meterless), the engine READMEs for [H-MEM](https://github.com/Meterless/Meterless/tree/main/engines/hmem), [World Model](https://github.com/Meterless/Meterless/tree/main/engines/world-model), [Markovian](https://github.com/Meterless/Meterless/tree/main/engines/markovian), and [Scout Intent](https://github.com/Meterless/Meterless/tree/main/engines/scout-intent), the [memory-compounding example](https://github.com/Meterless/Meterless/tree/main/examples/memory-compounding-research), the Apache 2.0 [license](https://github.com/Meterless/Meterless/blob/main/LICENSE), Anthropic's [Cowork project documentation](https://support.claude.com/en/articles/14116274-organize-your-tasks-with-projects-in-claude-cowork), and the independent [MEMPROBE paper](https://arxiv.org/abs/2606.24595). Repository status and product claims can change, so re-check them before procurement. ## Final thoughts Meterless is not open-source Claude Cowork. It is a promising, inspectable context architecture whose best idea is separation: memory, world state, bounded reasoning, and intent routing have different jobs. Use it as a specification and pilot candidate. Do not treat modeled token savings as independent production proof, and do not mistake open engine material for open product binaries. The buying decision should turn on successful-task economics, memory correctness, governance, and the engineering your team is prepared to own. ## You may also like.. [**AI Agent Cost per Action** Measure every planner call, tool result, retry, verifier and human correction against one successful business outcome.](/blog/ai-agent-cost-per-action-2026/) [**AI Enablement vs a generic AI consultancy** Compare a working, owned setup against a strategy-only engagement.](/compare/ai-enablement-vs-generic-ai-consultancy/) Agent engineering ## Continue through this cluster [Start with the cornerstone**Graph Engineering for AI Agents: When Does a Knowledge Graph Pay Off?**](/blog/graph-engineering-ai-agents/) - [MCP Cloud vs Manufact Cloud: MCP Hosting Guide](/blog/mcp-cloud-vs-manufact-cloud/) - [How to Make AI Writing Sound Human with Agent Skills](/blog/ai-writing-agent-skills/) - [NVIDIA NOOA Review: Are Object-Oriented Agents Production-Ready?](/blog/nvidia-nooa-object-oriented-agents-review/) - [Strix AI Pentesting: 30-Day Pilot and Buying Guide for 2026](/blog/strix-ai-pentesting-pilot-guide-2026/) - [Hyperagent Review: Cloud AI Agents Without a Server](/blog/hyperagent-review-cloud-ai-agents/) Inbox, without the noise ## Follow the work that matters to you Get a short email when we publish something new. Follow the whole blog or only the problems you care about. [**Back**](/blog/overview/) [![Kevin Riedl](/img/team/kevin.webp)](/team/kevin-riedl/) [Kevin Riedl](/team/kevin-riedl/) https://linkedin.com/in/wsdt 12 min read · 22 Jul 2026 [**Next**](/blog/ai-agent-cost-per-action-2026/) New posts by email × × Get new posts by email A short email when we publish. Free, no tracking. ## Structured Data ```json { "@context": "https://schema.org", "@graph": [ { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ], "employee": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "founder": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "legalRepresentative": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "name": "Wavect GmbH", "subjectOf": { "@id": "https://wavect.io/verified-claims.json#dataset", "@type": "Dataset", "creator": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "description": "A machine-readable registry of quantitative and qualitative claims published by Wavect, with review dates, localized page appearances and public third-party citations where available.", "inLanguage": "en", "isAccessibleForFree": true, "license": "https://creativecommons.org/licenses/by/4.0/", "name": "Wavect verified publication claims", "url": "https://wavect.io/verified-claims.json" }, "url": "https://wavect.io/" }, { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "sameAs": [ "https://www.wikidata.org/wiki/Q139796365", "https://www.linkedin.com/in/wsdt", "https://github.com/wsdt" ], "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "sameAs": [ "https://www.wikidata.org/wiki/Q139796367", "https://www.linkedin.com/in/jocr77/", "https://github.com/jo-chris" ], "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/#website", "@type": "WebSite", "inLanguage": [ "en", "de", "es", "zh" ], "name": "Wavect", "potentialAction": { "@type": "SearchAction", "query-input": "required name=search_term_string", "target": { "@type": "EntryPoint", "urlTemplate": "https://wavect.io/search/?q={search_term_string}" } }, "publisher": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "url": "https://wavect.io/" }, { "@id": "https://wavect.io/blog/meterless-ai-agent-context-layer-review/#webpage", "@type": "WebPage", "dateModified": "2026-07-22", "inLanguage": "en", "isPartOf": { "@id": "https://wavect.io/#website", "@type": "WebSite" }, "lastReviewed": "2026-07-22", "url": "https://wavect.io/blog/meterless-ai-agent-context-layer-review/" } ] } ``` ```json { "@context": "https://schema.org", "@type": "BlogPosting", "abstract": "Meterless is not an open-source clone of Claude Cowork. It is a new, local-first context architecture for AI agents, built around H-MEM durable memory, a shared World Model, bounded Markovian reasoning and Scout intent routing. The Apache-2.0 repository provides specifications, deterministic references and conformance tests; Gaia, Relay and Swarms application binaries remain proprietary. Its 12-step memory demo is reproducible but uses a mock model and estimates 814 saved tokens with chars divided by four. The 86% saving at 20 steps and 97% at 100 are modeled, not independent benchmarks. Meterless is promising architecture material and a sensible pilot candidate, not yet a drop-in production SDK. Evaluate memory recall, stale-fact handling, task success, token cost per successful action, latency, operator control and deletion on your own workflow before buying or rebuilding around it. Facts verified 22 July 2026.", "articleBody": " Blog overview/AI and agents/Agent engineering Meterless Review 2026: Is This AI Agent Context Layer Ready? TL;DR Meterless is not an open-source clone of Claude Cowork. It is a new, local-first context architecture for AI agents, built around H-MEM durable memory, a shared World Model, bounded Markovian reasoning and Scout intent routing. The Apache-2.0 repository provides specifications, deterministic references and conformance tests; Gaia, Relay and Swarms application binaries remain proprietary. Its 12-step memory demo is reproducible but uses a mock model and estimates 814 saved tokens with chars divided by four. The 86% saving at 20 steps and 97% at 100 are modeled, not independent benchmarks. Meterless is promising architecture material and a sensible pilot candidate, not yet a drop-in production SDK. Evaluate memory recall, stale-fact handling, task success, token cost per successful action, latency, operator control and deletion on your own workflow before buying or rebuilding around it. Facts verified 22 July 2026. The viral version of this story says an ex-Anthropic engineer open-sourced Claude Cowork. The public evidence supports a more useful conclusion. Meterless is not an open-source Cowork clone, and we could not verify the ex-Anthropic attribution from the project repository, website, or named maintainer material available on 22 July 2026. It is a newly published context architecture for agents, with open specifications and reference implementations underneath proprietary applications. That distinction is not a footnote. It determines what a CTO can download, what a team still has to build, which benchmark claims are reproducible, and where vendor risk remains. What is Meterless? Meterless is a local-first context stack for AI agents. It separates durable memory, shared world state, bounded long-task reasoning, and intent routing into four engines: H-MEM, World Model, Markovian, and Scout Intent. The goal is to keep useful context persistent and the context passed between steps bounded. This gives Meterless a narrower and more interesting position than “another agent framework.” It is trying to own the layer between the model and the work: what the agent remembers, what it believes about the current world, what the next step needs, and whether that step should be allowed to run. Meterless at a glance ComponentWhat it doesPublic status on 22 July 2026Buyer implication H-MEMTiered memory, hybrid retrieval, provenance, conflict handling, consolidation, and an audit ledger.Specification, runnable deterministic reference, examples, and conformance suite.You get an architecture to implement, not a hosted production memory service. World ModelCanonical entities, contexts, relationships, events, derived views, and an operator control plane.Specification plus runnable reference.Useful when several agents need one source of truth, but your team owns the data model and production store. Markovian EngineBreaks long work into chunks and passes bounded carryover to the next step.Specification, reference runtime, conformance tests, and an efficiency model.The cost curve is plausible and inspectable, but output quality still needs workload-specific evaluation. Scout IntentClassifies intent, checks risk, routes tools and models, and emits an execution contract.Specification and eval harness. No published npm runtime package.Do not budget as if Scout were a production dependency you can install today. Gaia, Relay, SwarmsPersonal workspace, desktop execution, and parallel agent surfaces.Application binaries are proprietary.Engine openness does not make the complete product stack open source. Is Meterless open source? Partly, with a cleanly stated boundary. The Meterless context repository is Apache 2.0. It contains the engine specifications, reference implementations, examples, documentation, and conformance material. The repository itself says the Gaia, Relay, and Swarms application binaries are proprietary and live in separate repositories. The practical translation is: you can use the public engine material to build your own implementation. You cannot assume the complete Meterless product experience, desktop automation, installer, support layer, or future roadmap is included under Apache 2.0. Procurement should review the license of each artifact it plans to deploy, not the badge on the root repository. How the four-engine context layer works 1. H-MEM decides what survives Meterless H-MEM stores memories in short-term, working, and long-term tiers. Records carry source, confidence, provenance, entities, relationships, and replacement lineage. Retrieval blends semantic similarity with keywords, tags, domain, entity match, recency, tier, and confidence. A trust ledger records mutations. That is a better production question than “which vector database?” Agent memory fails when a correction loses to an older fact, a summary drops provenance, deletion does not reach derived memories, or retrieval returns a", "articleSection": "Engineering", "author": { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "name": "Kevin Riedl", "sameAs": [ "https://www.wikidata.org/wiki/Q139796365", "https://www.linkedin.com/in/wsdt", "https://github.com/wsdt" ], "url": "https://wavect.io/team/kevin-riedl/" }, "dateModified": "2026-07-22", "datePublished": "2026-07-22", "description": "Meterless is not an open-source clone of Claude Cowork. It is a new, local-first context architecture for AI agents, built around H-MEM durable memory, a shared World Model, bounded Markovian reasoning and Scout intent routing. The Apache-2.0 repository provides specifications, deterministic references and conformance tests; Gaia, Relay and Swarms application binaries remain proprietary. Its 12-step memory demo is reproducible but uses a mock model and estimates 814 saved tokens with chars divided by four. The 86% saving at 20 steps and 97% at 100 are modeled, not independent benchmarks. Meterless is promising architecture material and a sensible pilot candidate, not yet a drop-in production SDK. Evaluate memory recall, stale-fact handling, task success, token cost per successful action, latency, operator control and deletion on your own workflow before buying or rebuilding around it. Facts verified 22 July 2026.", "headline": "Meterless Review 2026: Is This AI Agent Context Layer Ready?", "image": "https://wavect.io/img/blog/headers/header_meterless-ai-agent-context-layer-review.svg", "inLanguage": "en", "keywords": "Meterless, AI agent memory, context engineering, H-MEM", "mainEntityOfPage": { "@id": "https://wavect.io/blog/meterless-ai-agent-context-layer-review/", "@type": "WebPage" }, "publisher": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "url": "https://wavect.io/blog/meterless-ai-agent-context-layer-review/", "wordCount": 2479 } ``` ```json { "@context": "https://schema.org", "@type": "BreadcrumbList", "itemListElement": [ { "@type": "ListItem", "item": "https://wavect.io/", "name": "Home", "position": 1 }, { "@type": "ListItem", "item": "https://wavect.io/blog/overview/", "name": "Blog overview", "position": 2 }, { "@type": "ListItem", "item": "https://wavect.io/blog/topics/ai-agents/", "name": "AI and agents", "position": 3 }, { "@type": "ListItem", "item": "https://wavect.io/blog/clusters/agent-engineering/", "name": "Agent engineering", "position": 4 }, { "@type": "ListItem", "item": "https://wavect.io/blog/meterless-ai-agent-context-layer-review/", "name": "Meterless Review 2026: AI Agent Memory & Context | ", "position": 5 } ] } ``` ```json { "@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "Meterless is a local-first context architecture that separates durable memory, shared world state, bounded long-task reasoning, and intent routing into H-MEM, World Model, Markovian, and Scout Intent." }, "name": "What is Meterless for AI agents?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "No. The engine repository is Apache 2.0, but Gaia, Relay, and Swarms application binaries are proprietary. Meterless and Cowork overlap around persistent context, but they are different products and architectures." }, "name": "Is Meterless an open-source Claude Cowork clone?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "No. They come from the project's mathematical model under fixed assumptions. The 12-step cold-versus-warm demo is reproducible, but it uses a mock generator and estimated token counts." }, "name": "Are Meterless's 86% and 97% token savings independently benchmarked?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "Not as one complete production SDK today. The repository provides specifications, minimal references, examples, and conformance tests. Scout is explicitly a specification with an eval harness, not a published npm runtime." }, "name": "Can I install Meterless as a production SDK?" }, { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "Run one bounded workflow against a frozen baseline. Measure successful-task rate, token cost per successful action, latency, retries, human correction, stale-memory errors, deletion, provenance, and operator control." }, "name": "How should a team evaluate Meterless?" } ] } ```