{"slug": "show-hn-speck-cognitive-architecture-for-small-local-llms", "title": "Show HN: Speck – Cognitive architecture for small local LLMs", "summary": "Speck, a cognitive runtime built on the Genesis runtime, was released as a Show HN project that moves memory, planning, evidence tracking, confidence estimation and metacognition out of the prompt and into persistent deterministic software so that 1B–7B local language models can offload cognitive work. Speck stores cognitive state independently of the loaded model, letting a worker unload, swap or migrate models without losing task state, and it incorporates mechanisms derived from the Artificial Cognitive Architecture Omega Gen2 without being a fork or rewrite of it.", "body_md": "**Speck is a cognitive runtime designed to make small language models substantially more capable by moving cognition out of the prompt and into persistent, deterministic software.**\n\nInstead of repeatedly asking an LLM to simulate memory, attention, planning, evidence tracking, confidence, learning and metacognition, Speck implements those mechanisms in the runtime itself.\n\n**The model is disposable. The cognitive state is not.**\n\nA worker can unload its model, change models, restart, or migrate to another compute device without losing the cognitive state of its task.\n\nSpeck is built on the Genesis runtime and incorporates mechanisms derived from experiments in [Artificial Cognitive Architecture Omega Gen2](https://github.com/doctarock/Artificial-Cognitive-Architecture-Omega-Gen2), but it is **not a fork or rewrite of Omega**.\n\nMost LLM agents place an enormous amount of responsibility on the model.\n\nThe model is expected to:\n\n- remember what happened\n- determine what matters\n- maintain task state\n- recognise contradictions\n- track evidence\n- estimate confidence\n- plan\n- reconsider failed approaches\n- decide what to retain\n- use tools correctly\n- learn from outcomes\n- produce the final answer\n\nMuch of this work does not inherently require a language model.\n\nSpeck starts from a different principle:\n\n**Anything that can be remembered, measured, ranked, checked, persisted, scheduled, validated, learned procedurally or calculated deterministically should not consume LLM intelligence twice.**\n\nThe LLM is therefore treated as a specialised semantic processor inside a larger cognitive system rather than as the entire system.\n\nThis is particularly important for small local models.\n\nA large model can often compensate for weak agent architecture with raw intelligence.\n\nA 1B–7B model cannot.\n\nSpeck attempts to provide the missing machinery.\n\nAt a high level, Speck separates **cognitive state** from **model inference**.\n\n```\n                    USER / ENVIRONMENT\n                           │\n                           ▼\n                        INTAKE\n                           │\n              ┌────────────┴────────────┐\n              │                         │\n              ▼                         ▼\n           MEMORY                    EVIDENCE\n         ACTIVATION                  / CLAIMS\n              │                         │\n              └────────────┬────────────┘\n                           ▼\n                     COMPETITION\n                           │\n                           ▼\n                    WORKING CONTEXT\n                           │\n             ┌─────────────┼─────────────┐\n             │             │             │\n             ▼             ▼             ▼\n          PLANNER        WORKER     METACOGNITION\n             │             │             │\n             └─────────────┼─────────────┘\n                           ▼\n                       VALIDATION\n                           │\n                    ┌──────┴──────┐\n                    ▼             ▼\n                  REPLY          TOOLS\n                    │             │\n                    └──────┬──────┘\n                           ▼\n                  OUTCOME / EVIDENCE\n                           │\n                           ▼\n               MEMORY + BELIEF UPDATE\n                           │\n                           └──────► next cycle\n```\n\nThe model participates in cognition.\n\nIt does not own cognition.\n\nSpeck contains runtime mechanisms for several processes normally delegated to prompts or large models.\n\nMemory is stored independently of the currently loaded model.\n\nThe runtime supports mechanisms including:\n\n- persistent task state\n- memory admission\n- embedding-based retrieval\n- activation\n- associative edges\n- memory graphs\n- grounding\n- domain anchors\n- contextual recall\n\nRelevant memories can compete for inclusion in working context rather than simply being dumped into the model's context window.\n\nNot everything remembered should occupy the model's attention.\n\nSpeck can rank and compete candidate information before constructing working context.\n\nThis allows limited model context to be spent on information that is more likely to matter to the current task.\n\nSpeck distinguishes between information being present and information being established.\n\nThe runtime can maintain:\n\n- claims\n- supporting evidence\n- conflicting evidence\n- confidence\n- provisional hypotheses\n- belief revision\n\nNew evidence can therefore change persistent state without requiring the model to reconstruct the entire reasoning history.\n\nCandidate information can be checked against user statements and existing evidence before it becomes persistent knowledge.\n\nGrounding can combine semantic similarity with deterministic signals rather than relying entirely on an LLM to decide what the user previously said.\n\nUncertain information can remain task-local until sufficiently supported.\n\nPlanning is separated from execution.\n\nDifferent models can therefore be assigned to different cognitive roles.\n\nFor example:\n\n```\nIntake       → small fast model\nPlanner      → stronger reasoning model\nWorker       → general-purpose model\nValidator    → independent model or deterministic check\nEmbeddings   → dedicated embedding model\n```\n\nThere is no requirement for every role to use the same model.\n\nSpeck maintains runtime information about its own performance rather than merely prompting a model to \"reflect.\"\n\nThis includes mechanisms for:\n\n- assessment\n- calibration\n- competence tracking\n- prediction\n- outcome comparison\n\nThe objective is not introspection for its own sake.\n\nIt is to improve future decisions.\n\nRepeated successful behaviour should not require rediscovery forever.\n\nSpeck can represent procedures independently of conversational memory, allowing successful patterns to become reusable runtime knowledge.\n\nThe long-term goal is simple:\n\n**Do not repeatedly spend inference on problems the system has already learned how to solve.**\n\nSpeck is deliberately developed against small local models.\n\nThis exposes weaknesses that large frontier models can hide.\n\nSmall models may:\n\n- misunderstand complex prompts\n- lose state\n- mishandle negation\n- hallucinate structure\n- perform poor arithmetic\n- choose inappropriate tools\n- accept unsupported conclusions\n- prematurely ask the user for information\n- drift from supplied evidence\n\nSpeck does not assume these problems can all be solved with better prompting.\n\nWhere practical, responsibility is moved out of the model entirely.\n\n```\nLLM responsibility\n        │\n        ▼\nCan software perform this reliably?\n        │\n   ┌────┴────┐\n  YES        NO\n   │          │\nRuntime      Model\nmechanism   judgement\n```\n\nModel outputs can then be treated as proposals rather than unquestioned state transitions.\n\nOne of Speck's fundamental design constraints is:\n\n**No model should be the identity of the agent.**\n\nModels are replaceable cognitive resources.\n\nA task may begin using one worker model and continue using another.\n\nA model can be:\n\n- unloaded\n- upgraded\n- downgraded\n- replaced\n- moved to another device\n- assigned a different cognitive role\n\nwithout discarding the persistent state surrounding the task.\n\nThis makes heterogeneous local inference practical.\n\nSpeck includes a tool layer rather than allowing arbitrary model output to directly become action.\n\nThe runtime contains support for:\n\n- tool registration\n- tool intents\n- workspace operations\n- transactional workspace changes\n- browser automation\n- Genesis tools\n- validation before execution\n\nTool selection and tool execution can therefore be independently inspected and constrained.\n\nSpeck is built on the **Genesis runtime**.\n\nGenesis provides the surrounding agent infrastructure, including plugin hosting, profiles, UI integration and runtime services.\n\nSpeck provides the cognitive layer.\n\nConceptually:\n\n```\n┌────────────────────────────────────────────┐\n│                  SPECK                     │\n│                                            │\n│ Memory • Attention • Evidence • Beliefs   │\n│ Planning • Metacognition • Procedures     │\n│ Validation • Cognitive State              │\n└─────────────────────┬──────────────────────┘\n                      │\n┌─────────────────────▼──────────────────────┐\n│                 GENESIS                    │\n│                                            │\n│ Plugins • Tools • Profiles • Runtime      │\n│ Services • Interfaces • Infrastructure    │\n└────────────────────────────────────────────┘\n```\n\nSpeck is intended to be experimentally testable.\n\nThe repository includes benchmarking infrastructure for:\n\n- role qualification\n- trajectories\n- tool behaviour\n- validation\n- comparison\n- ablation testing\n- cognitive metrics\n\nThis is important because adding more cognitive machinery does not automatically make an agent better.\n\nA mechanism should be removable and testable.\n\nIf removing a component produces no measurable difference, its value should be questioned.\n\nA conventional agent often resembles:\n\n```\nPrompt\n  ↓\nLLM\n  ↓\nTool\n  ↓\nLLM\n  ↓\nTool\n  ↓\nLLM\n  ↓\nAnswer\n```\n\nSpeck is closer to:\n\n```\nEnvironment\n     ↓\nPersistent Cognitive State\n     ↓\nEvidence + Memory + Attention\n     ↓\nWorking Context\n     ↓\nSpecialised Model Judgement\n     ↓\nIndependent Validation\n     ↓\nAction\n     ↓\nMeasured Outcome\n     ↓\nBelief / Memory / Procedure Update\n     ↓\nNext Cognitive Cycle\n```\n\nThe distinction is intentional.\n\nSpeck is not primarily attempting to build a better prompt loop.\n\nIt is attempting to build the machinery surrounding the model.\n\nSpeck is designed to work with locally hosted models and services.\n\nA default installation:\n\n- binds to loopback\n- stores runtime data locally\n- does not require a cloud model\n- does not enable model, embedding or browser services unless configured\n- keeps credentials and personal runtime data outside the repository\n\nCloud models can still be used when desired.\n\nThey are resources available to the cognitive system rather than a requirement for the architecture.\n\n- Node.js 18 or newer\n\nClone the repository:\n\n```\ngit clone https://github.com/doctarock/Speck.git\ncd Speck\n```\n\nInstall dependencies:\n\n```\nnpm install\n```\n\nStart Speck:\n\n```\nnpm start\n```\n\nThen open:\n\n```\nhttp://127.0.0.1:4310\n```\n\nFresh runtime data will be created under:\n\n```\ndata/\n```\n\nModel, embedding and browser services remain disabled until explicitly configured.\n\nBuild:\n\n```\nnpm run build\n```\n\nType-check:\n\n```\nnpm run check\n```\n\nRun the test suite:\n\n```\nnpm test\n```\n\nRun the benchmark suite:\n\n```\nnpm run benchmark:suite\n```\n\nSpeck is an active experimental project.\n\nThe architecture is functional, but it should not yet be interpreted as a claim that every cognitive mechanism improves every task or every model.\n\nThe project deliberately includes benchmarks, probes and ablation infrastructure so those claims can be tested rather than assumed.\n\nSome mechanisms will work.\n\nSome will need refinement.\n\nSome may eventually be removed.\n\nThat is part of the experiment.\n\nSpeck incorporates ideas explored in **Artificial Cognitive Architecture Omega Gen2**, particularly around persistent cognition, memory, attention, competition, evidence and cognitive state.\n\nThe objectives are different.\n\n**Omega asks:**\n\nWhat happens if we attempt to construct increasingly mind-like persistent artificial cognition?\n\n**Speck asks:**\n\nHow much model intelligence can be replaced or amplified by persistent computational cognitive machinery?\n\nOmega explores artificial cognition.\n\nSpeck attempts to make that cognition useful as infrastructure.\n\n1. \n**The model is disposable. The cognitive state is not.**\n2. \n**Do not use inference for deterministic work.**\n3. \n**Models make semantic judgements; software enforces policy.**\n4. \n**Model output is evidence, not automatically truth.**\n5. \n**Memory should compete for attention rather than flood context.**\n6. \n**Successful reasoning should become reusable knowledge where possible.**\n7. \n**Failures should modify future behaviour.**\n8. \n**Different cognitive jobs may require different models.**\n9. \n**Small models are a constraint, not an afterthought.**\n10. \n**Cognitive mechanisms should be measurable and removable.**\n\nSpeck ultimately tests a fairly simple hypothesis:\n\n**How much of what we currently call LLM intelligence actually needs to live inside the LLM?**\n\nModern agents repeatedly ask enormous neural networks to remember, organise, reconsider, compare, track, schedule and validate information that conventional software can often handle more reliably.\n\nSpeck moves those responsibilities outward.\n\nIf successful, increasingly capable agents should become possible using smaller models, less inference, persistent knowledge and measurable cognitive machinery.\n\nThe goal is not to make a small model pretend to be a large model.\n\nThe goal is to give the small model a better brain around it.\n\n**Small model. Persistent mind.**\n\nSpeck is inspired by mechanisms developed in Artificial Cognitive Architecture Omega Gen2, but it is **not a rewrite or fork of Omega**.\n\nOmega is an experiment in artificial cognition and autonomous mind-like behaviour.\n\nSpeck has a different objective:\n\nUse conventional software to provide memory, attention, state, learning, planning support, metacognition, evidence tracking and proceduralisation so that a small LLM only performs operations that genuinely require semantic intelligence.\n\nThe central design principle is:\n\nAnything that can be remembered, measured, ranked, checked, persisted, scheduled, validated, learned procedurally or calculated deterministically should not consume LLM intelligence twice.\n\nA second critical principle:\n\nThe model is disposable. The cognitive state is not.\n\nA worker must survive unloading its model, changing models, restarting the process, or migrating to another compute device without losing its task state.\n\nThis distribution includes Speck's compiled server, browser interface, and the Genesis runtime modules it uses. It does not include optional plugins, credentials, or user data.\n\nRequirements: Node.js 18 or newer.\n\n```\nnpm install\nnpm start\n```\n\nOpen `http://127.0.0.1:4310`. The server binds to loopback by default and creates fresh runtime data under `data/`. Model, embedding, and browser services are disabled unless explicitly configured. Keep credentials and personal data in local settings, never in this repository.", "url": "https://wpnews.pro/news/show-hn-speck-cognitive-architecture-for-small-local-llms", "canonical_source": "https://github.com/doctarock/Speck", "published_at": "2026-10-06 08:50:01+00:00", "updated_at": "2026-10-06 09:19:54.062913+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-tools", "ai-infrastructure"], "entities": ["Speck", "Genesis runtime", "Artificial Cognitive Architecture Omega Gen2"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/show-hn-speck-cognitive-architecture-for-small-local-llms", "markdown": "https://wpnews.pro/news/show-hn-speck-cognitive-architecture-for-small-local-llms.md", "text": "https://wpnews.pro/news/show-hn-speck-cognitive-architecture-for-small-local-llms.txt", "jsonld": "https://wpnews.pro/news/show-hn-speck-cognitive-architecture-for-small-local-llms.jsonld"}}