Concepts

A program that decides for itself which tool to call next

Different Levels of AI Involvement in Software Development

An Architecture to Give AI Persistent Memory for Later Recall

Claude Code / Codex CLI / Gemini CLI and other AI Development Tools

A web framework with zero JS baseline

10 types of preferred attitudes (style/technical selection/design principles) + version chain + confidence decay

Claude Code leaked memory consolidation feature - four stages Orient → Gather → Consolidate → Prune, the naming origin of dream circles

Background track of memvault - persistent organization and forgetting after memory entry

Open-source 3D modelling and animation software

Cascade Recall: slow path along the map, L0/L1/L2 three layers KG

Customize zsh function, eat ANTHROPIC_BASE_URL + ANTHROPIC_MODEL env var, let tmux run different model per pane.

Community tool, change ~/.claude/settings.json domain-wide file cut provider, can't do tmux multi-pane and line different model so give up.

OpenAI's conversational AI

Anthropic official CLI agent, the main force of the planning layer of this architecture, the interface has been modified by CCR to run any model.

The extension mechanism added in Claude Code 2.1.287: small programs written by others can change how the terminal screen is drawn, without changing what the AI reads

Native :3456 protocol transformer, which represents Anthropic API for Claude Code, and forwards downstream to OpenAI.

Command Line Interface - a universal interface with zero schema overhead

Cognitive Load Theory - A Design Philosophy for Controlling Cognitive Load: Cutting Redundancy, Removing Complexity, and Calibrating Paradigms

OpenAI's official coding CLI, one of the organization's acceptance tiers, looks at the worktree git diff to provide a perspective independent of coding rules.

Engineers only point the way, development and testing are all handed over to AI.

Inline preprocessing layer for date/amount/ratio/length/Chinese punctuation standardization - 4 submodules + preprocess_chinese

Efficiency of context window usage and management policies for AI Agents

Token and API Cost Management for AI Tools

The ability to resurface memories across conversations and cycles.

Daily SQL aggregation out of top intents / entities, automatic diary writing

Engineers Just Set the Direction; Development and Testing Are Entirely Handled by AI

Dream Loops - Daily Batch Memory Integration Pipeline

A system for swapping a character's hairstyles, outfits and expressions

An AI text-to-speech and voice design service

Fault-tolerance strategy of letting rather than blocking a Handler crash.

Service → SDK → CLI → MCP → Skill Layered Policy

Animation drawn frame by frame, often on twos (each drawing held for two frames).

The execution framework that determines a model’s actual performance is more than just the model’s shell—using a different harness with the same model can result in a 10-percentage-point difference in scores.

Deterministic security guardrail mechanism for AI coding tools

Different Models of Division of Labor between Humans and AIs

Language Intent + BM25 Keyword Dual Indexing

7/30/90 Day Attention Window + Knowledge Gap Detection

Design Thinking Record from Conceptualization to Realization

Structured Data Markup - Enabling Search Engines to Understand the Semantics of Content

Repeated failures (verdict=INCORRECT, ≥2 failures) are labeled as knowledge gaps.

Extracting entities from conversations and building maps of relationships

Four Levels of Progressive Examination of Knowledge Graphs - Graph Structure, Ground Truth, LLM, Three Stages of Confirmation

The North Star of Memory Systems - Outdated, Noise, and Contradictions Must Retire

Making a character's mouth move with the audio

Multi-provider unified OpenAI-compatible gateway. Only model_list/key/fallback centralized in this architecture.

Dream Loop Phase 3 - Reading Statistical Snapshots to Generate Insights and Knowledge Gaps with LLM

Model Context Protocol - Standardized protocol for AI tools

Personal Memory Vault - Organize, Noise Reduce, De-Duplicate, and Save AI Conversations

MiniMax's video generation model, which accepts reference images and audio together.

Adobe's library of humanoid character animations

The system is broken down into independent and combinable modules.

Recording a real performer's full-body movement and applying it to a 3D character.

Models that take a character image and a reference video and make the character move like the video.

KG Side Weight Fusion 5 signals: co-occurrence + session overlap + Adamic-Adar + type affinity + semantic

Event Logging and Monitoring of AI Agent Behavior

Open source software development model

A record of architectural iterations from conceptualization to implementation

Integration of the Write, Background, and Recall tracks into a single system map perspective

Personalized routing over intent classification - attention profile (7/30/90) × historical intent

Blocking masquerade attacks: authority/role/markdown/temporal/base64 five categories - a sister component of InjectionGuard.

Intent Classification: Six Intents to Determine LayerPlan and ScoringConfig

Foundation Abstractions for Function Synthesis: Operator / Pipeline / Observable / Subject Quad (RxJS-inspired)

Memory Read Tracks: Idea Sorting, Evaluation Pipeline, Cascading Recall

Export sanitization, symmetric with InjectionGuard - query side + prefetch cache double filtering

Low confidence extraction into waiting area - auto-approve rules + human-in-loop fallback

Meta's model that estimates a person's 3D body and full-body joint angles from a single image.

Disposable tools - low-frequency operations are authored in the sandbox in real time

Three sequential gates before writing - noise, injection protection, de-duplication

A motion transfer model that drives a character image from 3D poses.

11-stage scoring: time decay, trust, PPR, semantics, de-emphasis

Encapsulate the underlying services as a middle tier of the programmatic interface.

ByteDance's video generation model, which accepts reference images and reference audio together.

Search Engine Optimization - Search Engine Optimization Techniques

Site Map - Tell the crawler what pages are available

Top level abstraction, toolset packaged as reusable AI skills

L1 Basic → L2 Proficient → L3 Creative - Tracking Skill Proficiency

VoiceAgentRAG-inspired 5-op pipeline predictive prefetching + admission control

Server-Side Rendering - Server-Side Rendering

A shot-by-shot plan of framing, camera position, action and timing.

Claude Code has a built-in delegation mechanism that allows you to assign haiku and other sub-models to run simple tasks, saving the main dialog quota.

Amplifying Personal Output Work Patterns with AI

Proactive exploration - cross-community bridge, indirect-strong intermediary, knowledge-gap gap

LLM 3-Way Referee (MERGE / SUPERSEDE / COEXIST) + Low Confidence Upgrade RLM in Old and New Factual Fights

mutation thinking / write-test separation / invariants / runtime regression / mock boundaries / draft is not a finished product - verification level alignment rules.

Planning to keep the Claude Code, split the worker into smaller models for cheaper, and cross the workflow paradigm with two models for acceptance.

A JavaScript library for rendering 3D in the browser

warm→cold→frozen LLM digest compression during degradation (≤400 chars) - downlink only, not uplink

Terminal multiplexer — many panes in one session, still running when you reconnect

Reduce LLM context token consumption

An AI service that generates 3D meshes and textures from images

Salesforce benchmark, found that the success rate of cutting-edge LLM to find hidden bugs in their own code is less than 50%, which is the basis of this paper's "Acceptance is not given to small models".

A file format for 3D avatars, built on glTF

Alibaba's Wan-family motion transfer model that animates a character image from a reference video.

A self-supervised speech recognition model

Memorizing the path to the vault - three gate plus dual-rail writes

Generating 3D models from images or text with AI

memvault self-driven backbone: Write Pipeline, Dream Loop, Read Pipeline, Slow Thinker