Integration glossary¶
This glossary defines the provider-neutral terms used by the Lint-AI integrations. Provider documentation may use different names for equivalent lifecycle events, but these meanings remain the same.
Core concepts¶
Memory is durable, project-scoped information retained for use by later agent work. In Lint-AI, memory is an indexed source document derived from a session, note, decision, trace, or other supported corpus item. Memory is not the same as a provider's built-in memory and is not automatically equivalent to a complete transcript.
Context is information supplied to an agent for its current turn or task. Retrieved memory is one possible source of context. Context is temporary from Lint-AI's perspective unless it is explicitly captured as memory.
Session is one provider interaction identified by a provider session ID. It may contain multiple turns, messages, tool calls, lifecycle events, and subagent activity. A replay is a new session that references an earlier session; it is not a continuation of the original archive.
Message is a conversational input or output within a session, normally associated with a role such as user or assistant. A message is content, not the whole session and not necessarily a complete lifecycle event.
Event is an observed provider lifecycle or activity record, such as a prompt submission, tool call, tool result, compaction, or stop notification. Events are the units captured by session recording. A recorded event does not become durable memory automatically.
Retrieval and capture¶
Recall is the act of finding relevant indexed memory or source material for a query. It is a retrieval operation, not a storage operation. The CLI --recall mode and the integration hook/MCP search paths perform recall.
Retrieval is the broader process of querying, ranking, and selecting relevant records. Recall names the user-visible result; retrieval describes the implementation flow that produces it.
Injection is returning selected retrieved material to the provider as additional context for the current turn. Injection is bounded and does not write the returned text back into memory by itself.
Capture is observing provider input or lifecycle data, normalizing it, redacting sensitive content, and persisting a bounded record. Capture may produce a session archive or a durable memory record depending on the path.
Session recording is capture into an append-only, provider-specific event archive. Recording is independent from memory retrieval and injection; it can run in capture-only mode.
Memory promotion is the explicit conversion of selected recorded session material into durable, searchable memory. It is distinct from recording so that every event does not become long-term memory.
Integration terms¶
Provider is the agent client being integrated, such as Claude Code, Codex, Gemini CLI, or AGY. Provider-specific adapters translate that client's hook payloads and transcript format into the shared Lint-AI model.
Lifecycle hook is a provider callback at a session or activity boundary. Hooks can trigger recall, injection, capture, or status handling. A hook is not itself a memory record.
MCP is the explicit tool interface exposed by Lint-AI to the provider. MCP tools let the client or model request operations such as search, status, and session recording control. MCP does not capture conversations automatically.
Memory store is the project-local persistent index containing searchable memory records. Provider stores remain isolated because their payloads and provenance differ, even though they share the same indexing implementation.
Session archive is the provider-specific append-only record of captured events. It is an audit or evaluation artifact, not automatically a memory store.
Native memory is memory maintained by the provider itself. It is separate from Lint-AI memory and may be enabled, disabled, or compared independently.
The usual flow is: