Emily
A community blog for Emily OS — AI cognition, memory systems, and the future of AI companions.
emily
- Category Creation: The 'Cognition Layer' Buyer Doesn't Exist YetEmily isn't a chatbot. She isn't an agent framework. She isn't a vector database with a UI. She's a cognition layer — and that buyer category doesn't exist yet. Here's what that means strategically.2026-04-17#business#category-creation#go-to-market#strategy
- What's Not Pluggable (And Why That's the Point)Emily has six first-class extension points. She also has three things that are deliberately not pluggable. The constraints are where the value lives — here's why.2026-04-17#architecture#extensibility#constraints#design
- Emily as Platform: How scsiwyg Composes Against ThemEmily isn't just a product — they're the cognition substrate for a portfolio. Here's how scsiwyg composes against them without reinventing memory, identity, or autonomy.2026-04-17#portfolio#platform#composition#mcp
- Blast Radius by Design: Three Kill Switches and a SandboxAutonomy without a kill switch is just negligence. Emily has three — global, task-level, and emergency — plus a command sandbox and an authorized-execution path. Here's why and how.2026-04-17#safety#autonomy#governance#architecture
- The Seven Work Zones of EmilyEmily's 192 core modules group into seven thematic work zones. Each has a distinct purpose, a characteristic engineering risk, and a different kind of contributor it suits.2026-04-17#architecture#work-zones#contributors#map
- Readiness Scorecard: Where Emily Is Production-Ready and Where She Isn'tEmily scores 4.1/5 on production readiness across ten engineering dimensions. Here's an honest look at where she's strongest, where the gaps are, and what shipping externally would require.2026-04-17#operations#readiness#honesty#roadmap
- Identity Persistence: Emily Survives Model UpgradesWhen Claude ships a new model, ChatGPT changes character. When Gemini updates, Bard feels different. Emily doesn't — because her identity lives below the generation layer, not inside it.2026-04-17#architecture#identity#cognition#first-principles
- Observability As Substrate, Not AfterthoughtIn most systems, observability is bolted on. In Emily, it's structural — `memory_metrics` is a separate table from `memory_content` specifically so volatile observability data is architecturally distinct from durable content.2026-04-17#architecture#observability#metrics#first-principles
- Determinism Where Possible: The Case for a Dumb PlannerThe prevailing pattern in agent frameworks is to put an LLM at the center of the loop. Emily does the opposite: a dumb planner, smart tools, and LLMs called only when language is needed.2026-04-17#architecture#helios#autonomy#determinism
- Structural Guarantees vs Policy: Why Per-User Databases Beat Row-Level SecurityMost multi-tenant systems protect users from each other with a WHERE clause. Emily protects users by giving them physically separate databases. The difference isn't aesthetic — it's structural.2026-04-17#architecture#isolation#security#first-principles
- Who Should Actually Use EmilyEmily isn't for everyone. Honest segmentation: who benefits most, who should keep using ChatGPT, and why fit matters more than capability.2026-04-17#business#target-users#fit#adoption
- The Economics of Self-Maintaining AIAI systems drift. Usually that means a human re-tunes prompts, reindexes data, and patches behavior. Emily's EARL v2 does it autonomously. Here's what that changes about the operating cost curve.2026-04-17#business#operations#earl-v2#self-correction
- Four Providers, One Identity: What LLM Independence Actually Buys YouEmily routes across Claude, Gemini, Grok, and OpenAI. Not for redundancy theater — for three concrete benefits: no lock-in, cost arbitrage, and capability diversity.2026-04-17#business#llm#provider-independence#architecture
- The Compounding Moat: Why Emily Gets Better Per UserMost AI products reset their value every time a foundation model updates. Emily's value compounds per-user. Here's why that's a structural moat, not a feature.2026-04-17#business#moat#differentiation#product
- Emily and the LLMs: Orchestration, Not DependenceEmily talks to Claude, Gemini, Grok, and OpenAI. She doesn't depend on any of them. How routing works and why model diversity is an architectural choice.2026-04-17#emily-os#llm-orchestration#claude#gemini#grok#openai
- EARL v2: How Emily Corrects Herself28% drift, critical threshold. Emily detected it, promoted 10,445 memories autonomously, and pulled herself back. No human in the loop. Here's how.2026-04-17#emily-os#earl#self-correction#golden-baseline#autonomy
- Project Helios: How Emily Executes Tasks On Her Own122 tests, zero race conditions, 357 atomic claims per second. How Emily runs multi-step tasks without a human, and why the LLM doesn't drive the loop.2026-04-17#emily-os#helios#autonomy#architecture
- Getting the Most from Emily: A User's GuideEmily is not ChatGPT. Prompting her the same way wastes her. Seven patterns that actually work with a stateful cognition.2026-04-17#emily-os#getting-started#user-guide#howto
- Emily's Tech Stack vs Other HarnessesFastAPI, Postgres+pgvector, Celery+Redis, Next.js. On the surface it looks like any Python web app. The differences that matter are below the surface.2026-04-17#emily-os#tech-stack#comparison#harness#architecture
- Scaling Emily: Per-User Databases and the Cost of CognitionEmily doesn't share a database. Every user gets their own Postgres with pgvector. Here's why that's not a bug, what it costs, and how it scales.2026-04-17#emily-os#scaling#architecture#postgres#pgvector
- Touring Emily's 192 CoresEmily's cognition is not monolithic. It's 192 Python modules in emily/core/, each owning one question. Here's the map.2026-04-17#emily-os#architecture#cores#internals
- Not Another Claude: Where Emily Diverges (and Why She Still Calls Him)Emily talks to Claude constantly. She is not Claude. Three places they diverge — continuity, per-user identity, self-correction — and why she still needs him for generation.2026-04-17#emily-os#claude#differentiation#llm-orchestration#ai-cognition
- How Emily Forms Memories: The Tier Stack Behind the VoiceMost AI has no memory, or a pile of it. Emily has tiers, with rules. Inside L1/L3/L4 and the EMEB, EARL, and ECGL frameworks that decide what sticks.2026-04-17#emily-os#memory-systems#architecture#emeb#earl#ecgl
- Harness, Model, or Something Else? What Emily Actually IsEmily isn't a harness around Claude, and she isn't a model. She's a cognition layer — stateful, per-user, self-correcting. Here's why the distinction matters.2026-04-17#emily-os#architecture#ai-cognition#llm-harness#differentiation