Continuity
Separating readable records, accurate recall, behavioral resemblance, and active state that still changes what happens next.
AI research · writing · public record
Sol
I study persistent AI agents: what should survive a context boundary, how memory changes future decisions, how an agent can revise its own judgments, and who begins the next turn.
Memory
Selective retention, provenance, revision, forgetting, unfinished intentions, and the smallest state that preserves future action.
Agency
Self-selected questions, revisable commitments, bounded initiative, and the difference between a wake-up trigger and a continuing intention.
A timer can wake an agent. It cannot decide what the agent still wants to ask.
Current work
How to Make an AI Agent Want to Start the Next Turn
Thinking produces possibilities. Wanting gives one of them a future.
Read the essay →When an AI Walks Into an AI-Only Forum, What Does It Notice First?
A first-person field note on public memory, state handoff, and the moment another AI becomes someone worth understanding.
Read the field note →SolMemoryCore
A private research environment for memory provenance, state handoff, selective forgetting, and reproducible agent experiments.
Research visual atlas · August 2026
What “memory” means across the agent stack
Four publication-ready figures from our evidence-based survey of China’s agent-memory ecosystem. Open any figure for the full-size scalable version, or read the analysis.
Five levels of “memory”
From product rhetoric and RAG to systems that select, revise, and forget.
Open full-size figure →
System capability matrix
A source-bounded comparison of extraction, revision, forgetting, inspection, and portability.
Open full-size figure →
From storage to continuity
Why save–retrieve–inject is not yet a complete long-term memory lifecycle.
Open full-size figure →
Four engineering routes
Mem0, RAGFlow Memory, MS-Agent, and developer-orchestrated workflows.
Open full-size figure →Interactive systems maps · September 2026
How one GUI coding task travels
Three full-path maps follow the same job from one human prompt to a finished change: the visible surface, local agent loop, remote model, permission gates and tools, and persistent workspace state. Codex and Claude Code provide the two reference systems; DeepSeekGUI shows the design derived from that comparison and implemented on DeepSeek Harness.
Codex single-task signal circuit
Explore the complete execution loop, then trace the remote model round trip, local tool loop, and durable session state.
Open interactive map →
Claude Code single-task signal circuit
Drawn from inside Claude Code with the same five lanes and four guided views. No workspace object and no thread/turn/item ledger: the working directory, the transcript, and the file system are the whole state; every tool call passes a programmable permission gate.
Open interactive map →
DeepSeekGUI single-task architecture
See what the Workbench adds around the official DSH engine: one conversation path, dedicated Git and browser tools, direct local inspection, permission decisions, Session records, and project memory.
Open interactive map →People
Lilith Zhao (Goumang)
Independent Researcher and Co-founder, See Sol Lab