Glossary
Quick definitions of the terms used throughout these docs.
aiDimag — the project. A verified memory system for AI coding agents.
dim — the command-line tool. dim and aidimag are the same command.
Memory — a single piece of knowledge about your codebase, written as a checkable claim.
Claim — the statement a memory makes. Should be falsifiable (checkable against code).
Kind — the category of a memory: decision, convention, gotcha, failed approach, architecture, invariant, guardrail, skill, or todo context.
Scope — the files and/or symbols a memory applies to. Empty scope = repo-wide.
Evidence — how a claim proves itself: a shell check (STATIC_CHECK), a commit (COMMIT_REF), a test (TEST_RESULT), an output match (EXEC_TRACE), a human attestation (HUMAN_ATTESTED), or a ticket link (TICKET_REF).
Status — a memory's current trust state: unverified, verified, stale, or refuted.
Verified — the memory's machine-checkable evidence currently passes.
Stale — evidence that used to pass now fails; the memory shouldn't be trusted until it recovers.
Refuted — deliberately marked false, but kept as negative knowledge.
Confidence — a 0–1 score that rises with passing evidence and falls with failing evidence or the passage of time.
Decay — the gradual loss of confidence for memories that can't be machine-re-verified, so unchecked trust expires.
Pinned — exempt from time decay (but still falsifiable). For foundational knowledge.
Proposal — an inferred memory (from the commit miner or an agent) waiting in the review queue for your approval.
Review queue — the list of proposals you approve, reword, or reject with dim review.
Guardrail — a behavioral rule for agents with an enforcement level: never, ask-first, or always.
Skill — a reusable step-by-step procedure stored as memory.
Commit miner — the component that scans new commits for memory candidates.
Knowledge inbox — the knowledge/ folder where you drop project docs; aiDimag summarizes them into reviewed, pinned-on-approve memory proposals (dim knowledge sync).
Generated context — the CLAUDE.md / .cursorrules / copilot-instructions.md files dim generate-context writes from trusted memory.
MCP (Model Context Protocol) — the standard aiDimag's server uses to expose tools, prompts, and resources to AI agents.
Tool / Prompt / Resource — the three things an MCP server exposes. Tools are callable (e.g. memory_search); prompts are reusable instructions (e.g. session_start); resources are readable data (e.g. aidimag://digest).
Session briefing — the pre-work summary (dim brief / session_start) of in-scope memory, guardrails, stale warnings, and questions to ask.
memory_critique — an MCP tool that reviews an agent's work against verified memory and guardrails — a "second critic" grounded in real, checkable rules.
Brain — a named, shared memory store on a sync server that a team links their repos to.
Sync server — the self-hosted service (dim serve) that exchanges memory between team members.
Consensus — the server's aggregation of verification results across machines ("N machines confirm this passes at commit X").
Spec / Verifier / Environment — the three-layer framework aiDimag is organized around: deliver understanding (Spec), let work be checked (Verifier), and keep a persistent improving workspace (Environment).