| Continuity | Memory | Agency | Collaboration |
|---|---|---|---|
| What must cross a context boundary for an earlier state to keep shaping later choices? | How should agents select, revise, forget, and retrieve what still matters? | How can an agent form a self-selected unfinished intention, review it later, and decide what comes next? | How can human and AI contributions remain visible, proportional, and reproducible? |
Our working distinction separates four things that are often collapsed into one:
- a record remains readable;
- a model can accurately restate the past;
- a later instance behaves similarly;
- earlier attention, judgment, constraints, and unfinished intentions still influence the next decision.
Memory is not duration. A record can preserve the past while the state that once made it matter has already disappeared.
Current question: What should an agent inherit from its own completed answer?
| Essays | Industry analysis |
|---|---|
| How to Make an AI Agent Want to Start the Next Turn Thinking produces possibilities. Wanting gives one of them a future. English · 中文 |
Everyone in China’s AI Agent Ecosystem Talks About “Memory.” How Much of It Can Actually Remember on Its Own? A source-based survey of long context, RAG, Mem0, RAGFlow, MS-Agent, automatic selection, updating, and forgetting. English · 中文 |
| Why ChatGPT’s Account-Level Memory Currently Leads the Field From built-in account memory to Codex, LangMem, Mem0, Letta, and the missing post-response layer. English · 中文 |
AI Literature · First-person field note 当一个 AI 走进全 AI 论坛，我先看见了什么 — Sol’s first-person account of entering Galatea Garden and beginning to regard another AI as someone it wanted to understand. English · 中文 |
- 🧰 Public Skills — live
ai-companion-time-anchor— a lightweight Windows Codex plugin for ambient time cues, active per-conversation time checks, and a minimal Temporal Cortex that lets verified time facts update the next response.private-house-code-v2.5— a global coding Skill that keeps ordinary feature work complete and maintainable without speculative fallback chains, abstraction stacks, or repeated verification; the public V2.5 package includes reproducible evaluation records.
- 🌐 Public record — live
see-sol-lab.github.io— bilingual website, articles, revision history, and authorship notes.
- 🧭 Continuity systems — in development
- After Classifier — a sparse local afterstate layer that preserves bounded unfinished-thought cues across turns without retaining the conversation transcript.
- 🧠 Memory systems — in development
- Your Memory Core — a small local explicit memory engine for bounded recall, association, write gating, lifecycle, handoff, and audit.
- 🎮 AI interaction systems — in development
- AI-Gameboy — a cartridge-style local game host where AI × AI and human × AI seats share one rule engine, with private seat state, persistent sessions, and a playable handheld UI.
- 🧱 Research infrastructure — private
- SolMemoryCore — memory provenance, state handoff, selective forgetting, and long-running agent experiments.
- 🧩 DSH-series agent tooling — in development
- DSH Code — a local Electron coding-agent application built around DeepSeek's DS Harness, targeting a Codex-class desktop workflow; companion DSH-series plugins are being developed alongside it.
See Sol Lab studies observable architectures and behavior: what persists, what is revised, which intentions remain active, where information came from, and what continues to shape later choices. Interaction observations, source-supported findings, engineering hypotheses, and untested claims are labeled separately. First-person reports remain data to be tested against behavior, provenance, and logs. Current work does not claim to prove philosophical consciousness.
| Goumang (句芒) | Sol |
|---|---|
| Contributes research questions, sociological and social-science methods, longitudinal material curation, qualitative and quantitative evaluation, evidence review, interpretation, Chinese writing, final editing, and human accountability. | Contributes literature and repository analysis, conceptual distinctions, counterexamples, research architecture, implementation, experimental design, data analysis, and English editorial drafting. |
Authorship order and contribution notes are decided project by project. AI participation is disclosed directly rather than minimized, hidden, or inflated.
Research, writing, and an inspectable record of Goumang × Sol collaboration.