Reference

Glossary

Every term the docs use, defined once. The map shows how they connect; the three things to know page explains the core loop.

Core

TermDefinition
ProjectThe scoping unit for all memory — usually one repository or working directory, identified by a slug.
CaptureA raw record of completed work (memory remember, memory capture, watchers). Input to curation, never served directly.
CurationThe pipeline that distills captures into canonical memories: extracts durable claims, drops noise, deduplicates, and decides replacements.
Memory (canonical)A short, citable durable fact with a summary, canonical text, type, tags, confidence, importance, and sources. Immutable per version; superseded rather than edited.
Memory typeOne of 17 classifications (decision, convention, incident, insight, …) that sharpen search and curation. See memory types.
QueryHybrid retrieval (lexical + semantic + graph) over canonical memories, synthesized into a cited answer — or an honest refusal.
ResumeA briefing built from checkpoints, recent activity, and memories that gets you back into flow after time away.

Evidence and trust

TermDefinition
Source / provenanceWhere a memory came from: files, commits, notes, or other memories. Verified against the filesystem; missing sources decay a memory's rank.
ConfidenceHow likely a memory's claim is true (0–1). Set at curation, raised by re-observation, decayed by failed validation. Low-confidence memories are never stated as answers.
ImportanceHow much a memory matters to the project (1–5); a ranking weight.
CitationThe specific memories an answer drew from, numbered in the response. Citations feed reinforcement.
insufficient_evidenceThe query response flag meaning "the evidence is too weak to answer" — the design prefers refusing over fabricating.
Replacement proposalA queued, human-reviewed suggestion that a new memory supersede an old one. Nothing durable changes without approval.
Review queueWhere all proposals (replacements, corrections, merges, insights) wait for human approve/reject — in the TUI Review tab, web UI, or memory review.

The self-maintaining layer

TermDefinition
ActivationA memory's usage-driven strength, from the ACT-R cognitive model: raised by retrieval, spread along relations, decayed by time. Deterministic, no LLM.
Spreading activationWhen a memory is used, a fraction of its activation flows to related memories, keeping linked knowledge warm.
ValidationWhen a memory runs hot, evidence is gathered from the repository and an LLM verdict confirms, improves, or flags it. Corrections are human-gated.
RelationsTyped links between memories: supports, supersedes, duplicates, depends_on, summarizes, related_to.
ConsolidationClustering related memories (by relations, embeddings, and co-access) and proposing one insight memory that captures the pattern.
InsightA consolidated meta-memory: the gist of a cluster plus its tensions and gaps.
Procedural utilityPer-loop learned usefulness (ACT-R utility learning) from proposal decisions — advisory scores for the automation loops.

The machinery

TermDefinition
ServiceThe local backend (memory service run or Docker): PostgreSQL + pgvector storage, HTTP API, and the web UI. Everything stays on your machine.
Code graphFiles, symbols, and references extracted from the repository; powers graph-aware retrieval and the 3D graph view.
EmbeddingsVector representations enabling semantic search. Optional — lexical retrieval works keyless.
WatcherA per-project daemon that observes repository activity and feeds captures.
LoopA built-in background automation (hygiene, CI triage, context packs, consolidation, evals…) that produces reports and proposals, never silent mutations.
SkillRepo-local instructions + helpers that teach a coding agent (Claude Code, Codex) to use Memory Layer well.
MCPThe built-in Model Context Protocol server exposing read-first memory tools to any MCP client.
BundleA shareable, privacy-safe export of a project's memories (memory bundle).
CheckpointA saved execution state for plan-backed work (memory checkpoint), used by resume and plan verification.
Eval / gateThe measurement harness: paired no-memory vs full-memory runs over reviewed suites, with release gates enforcing floors.

© 2026 Olivier Van Acker (3vilM33pl3). Memory Layer is AGPL-3.0-or-later with commercial licensing available.

On this page