Reference
Glossary
ACT-R — Adaptive Control of Thought—Rational, a cognitive architecture. Sigil uses its base-level learning equation to re-rank search results by recency × frequency, bounded to a (1 + 0.10 × activation) multiplier so it acts as a tiebreaker rather than overriding relevance. See Ranking.
Activation — a fact’s ACT-R score. Rises with access frequency and recency; decays as -0.5 × ln(t) in days, floored at one day.
AUDM — Add / Update / Delete / Merge. The decision every write goes through, comparing the incoming fact against existing memory. What stops accumulation. See Deduplication.
Agent scope — a visibility setting restricting a fact to the agent that wrote it. For instructions addressed to one assistant.
Category — the classification assigned to a fact, from a fixed vocabulary of fourteen: four personal (preference, opinion, personal, experience) and ten knowledge (business_rule, workflow, architecture, convention, decision, domain_knowledge, key_insight, metric, issue, action_item).
Chunk — a ~512-token block of an ingested document, contextualized and embedded. The raw material layer beneath facts.
Contextualization — the generated preamble attached to each chunk situating it in its document. Without it, “this is capped at 30 days” is unretrievable.
Confidence — the extractor’s certainty in a fact: high, medium, or low.
Daemon — the long-lived background process that holds the connection pool, caches, write queue, and peer endpoint. One per machine. sigil daemon status.
Device scope — a visibility setting restricting a fact to the machine that wrote it. For local paths, ports, hardware.
Directive — a virtual pod kind holding standing instructions, injected regardless of what the current prompt is about. Selected by category. See Pod kinds.
Embedding — a 1024-dimension vector representing text meaning. The dimension is a build-time constant, pinned in both the schema and every provider so the two can never drift.
Entity — a typed node in the knowledge graph: person, project, service, concept. Extracted from facts.
Entity graph — typed entities connected by extracted relations, traversed with recursive CTEs.
Fact — one atomic, self-contained statement. Sigil’s primary retrieval unit — what gets ranked, superseded, and injected.
Fresh — the lifecycle stage of a newly captured fact. Not yet in the canonical ranking pool. Promotes to stable after an hour, given vital importance or at least one access.
Hebbian edge — a strengthening link between two facts (or two entities) retrieved together. Neurons that fire together, wire together. Entity edges decay with a 30-day half-life.
Hook — a Claude Code extension point. Sigil registers four: UserPromptSubmit, PostToolUse, Stop, SessionEnd.
Hot context — the top-20 fact snapshot in ~/.sigil/CLAUDE.md, loaded on every session with zero latency. The query-independent recall channel. See Hot context.
Importance — vital or supplementary. Vital facts take global hot-context slots and survive lifecycle promotion without ever being accessed.
Ingest — storing a whole document: chunk, contextualize, embed, extract facts, link entities. Keeps the original text.
Injection floor — memory.injectionFloor, default 0.6. The relevance threshold for automatic hook injection. Deliberately higher than the search floor.
Iroh — the peer-to-peer transport used for device pairing.
Lifecycle stage — fresh, stable, or editing. See Memory decay.
MCP — Model Context Protocol. The standard Sigil exposes its thirteen tools over.
Managed session — an opt-in engine keeping warm claude CLI workers in tmux, amortizing agentic cold-start. Off by default.
Namespace — a hard partition. Facts in one namespace are invisible from another. A retrieval boundary, not a security boundary.
PGlite — Postgres compiled to WebAssembly, running in-process. Sigil’s embedded database mode. Zero install, single machine only.
pgvector — the Postgres extension providing vector similarity search. Required for external Postgres.
Playbook — a pod kind for procedural memory: runbooks, debug recipes. Active when its project matches yours.
Pod — a typed container for facts, whose type is a contract declaring decay, hot-context budget, retrieval weights, and active mode. Six built-in kinds. See Pods.
Pod kind — the contract. claude_session, project, person, playbook, vital, directive.
Prime — the MCP tool that loads hot context at session start. The equivalent of Claude Code’s CLAUDE.md import, for clients without hooks.
Provenance — the record of where a fact came from: source document, agent, device, timestamps, supersession links.
Relation — a typed edge between two entities in the graph.
RRF — Reciprocal Rank Fusion. Combines two ranked lists using rank alone rather than incomparable scores. K=20 in Sigil; vector weight 1.0, keyword weight 0.7.
Secret masking — pattern-based redaction at the single ingest choke point, before classification, chunking, embedding, and extraction. Runs on the read side too.
Session pod — a claude_session pod. One per Claude Code session, 90-day TTL.
Shim — a stable launcher script in ~/.sigil/bin/ that resolves the real binary at runtime. Every agent integration points at a shim, never a versioned path. sigil connect re-pins them.
Stable — the lifecycle stage of the canonical fact pool that retrieval ranks against.
Supersession — when a new fact contradicts an old one, the old one is marked superseded and linked to its replacement. It stops being retrieved but is not deleted, so provenance and point-in-time queries survive.
Synthesis — having an LLM write an answer from retrieved facts and chunks, rather than returning a ranked list.
tsvector — Postgres full-text search. Sigil’s keyword half of hybrid search, BM25-scored.
TTL — time to live. Declared per pod kind. Only claude_session has one (90 days).
Visibility — shared, agent, or device. Which agents and machines may retrieve a fact. sigil visibility <id> <scope>.
Vital — the highest importance level, and the virtual pod kind that surfaces such facts globally in six always-on hot-context slots.
Write policy — origin-only or open. Whether installs other than the pod’s creator may write to it. Matters in shared-database setups.