Home
Language: English · 简体中文
LWC is an agent-driven proactive memory CLI. It helps AI agents recall, maintain, and evolve persistent, source-grounded knowledge across sessions.
Instead of rebuilding an answer from raw document chunks every time, an agent can use LWC to maintain a durable Wiki of sources, pages, citations, links, decisions, and history. The Wiki gets more useful as the agent integrates new evidence and preserves verified conclusions.
- Project overview — what LWC solves, who it is for, and what it deliberately does not do.
- Basic concepts — sources, Wiki pages, provenance, scopes, graphs, Work, and changesets.
- Use cases — when persistent Agent memory pays off and when a simpler tool is enough.
- Installation and upgrades — supported installation paths and Agent setup.
- Quick start — initialize a project, integrate a source, recall knowledge, and verify the result.
| Capability | Purpose |
|---|---|
| Persistent Wiki | Keeps maintained knowledge instead of disposable chat answers. |
| Source grounding | Preserves immutable source snapshots, citations, provenance, and contradictions. |
| Bounded recall | Searches compiled pages first and opens exact source evidence only when needed. |
| Document knowledge graph | Traverses page, source, citation, link, and explicit semantic relationships. |
| CodeGraph | Answers structural questions about symbols, callers, callees, dependencies, and impact. |
| Word Graph | Explores shared terms across a bounded document sample without loading the entire token graph. |
| Strong tags | Loads a small, explicit set of complete core pages without relying on search. |
| Agent integration | Installs the LWC MCP, Skill, Hooks, and Instructions through host-specific AgentTarget adapters. |
| Safe change lifecycle | Uses changesets, checkpoints, durable Work, lint, and graph verification for recoverable maintenance. |
- Agent-operated: LWC is a machine interface for agents. People choose sources, goals, and review boundaries; agents handle the bookkeeping.
- Local-first: SQLite is canonical. Markdown and optional graph stores are rebuildable local projections.
- Source-grounded: Evidence remains immutable, while maintained pages can be revised as knowledge changes.
- Explicitly scoped: Project memory stays with the project. Reusable cross-project knowledge belongs in global memory.
- Auditable: Successful commands return JSON on standard output; failures return structured JSON on standard error.
- Consent-aware: Optional graph initialization and external integrations are never silently enabled.
Next: Project overview