Helix — Glossary

Shared vocabulary for the project. If a term here conflicts with code, the code is wrong — fix it (spec-first invariant, CLAUDE.md).

Term	Definition
Helix	The product: a local-first, portable, git-like memory layer for AI coding agents.
Strand	A user's whole memory as a single self-describing SQLite database; exported as a .dna file.
.dna	The portable strand artifact: signed (Ed25519), encrypted (XChaCha20-Poly1305), versioned.
Memory / Fact	One typed node in the graph (identity, preference, project, decision, entity, convention, snippet, fact).
Edge / Relation	A typed, weighted link between memories (e.g. has_decision, depends_on, supersedes).
Scope	Where a fact applies: global or project:<id>. Recall is scope-aware.
Ingestion	Accepting a user-routed slice (turn/code/note) for possible learning. No ambient capture.
Slice	A unit of routed input (a conversation turn, code block, or note) with metadata.
Redaction	Scrubbing secrets from a slice before any storage or model call (an invariant).
Heuristic gate	The cheap local check that decides whether a slice is worth extracting from — the main cost lever.
Extraction	Turning a slice into candidate typed facts (deterministic or LLM-backed).
Embedding	A vector representation of text used for semantic recall (local bge-small by default).
Consolidation	Deciding ADD / UPDATE / DELETE / NOOP for each candidate fact against existing memory.
Conflict resolution	Reconciling contradictory facts (recency > confidence > provenance; optional LLM tie-break).
Recall	Retrieving relevant memories for an agent (hybrid vector + graph + ranking).
Salience	A memory's current importance; decays over time unless reinforced.
Confidence	How sure Helix is a fact is true/durable.
Provenance	The record of what produced/changed a fact (slice, extractor, model, op) — the "why it believes this".
Decay / Reinforcement	Salience falling over time vs. rising when a fact is recalled/confirmed/re-seen.
Transfer	Moving memory between machines/agents/people via export/import/merge.
Merge	Combining two strands with consolidation + conflict resolution; reversible.
Rollback	Restoring a prior strand version from history.
Manifest	The plaintext, signed metadata of a .dna (schema, embedding space, counts, integrity root).
MCP	Model Context Protocol — the open interface through which agents read/write Helix memory.
LLM router	The component that picks/falls back across LLMs (free-tier-first) — or none at all.
Deterministic extractor	The no-LLM, rules+embeddings extractor; the $0 floor.
Daemon	The local Helix process hosting the engine for CLI, MCP server, and dashboard.
ADR	Architecture Decision Record — an entry in DECISIONS.md.
Episodic / semantic / procedural	The three long-term memory shapes Helix models (events / durable facts / skills), plus a cross-cutting entity graph.
Working memory	The live agent context window — used, but never Helix's system of record.
CLS (Complementary Learning Systems)	The fast-episodic-trains-slow-semantic model behind the two-stage write path.
Bi-temporal	Every fact carries valid-time (true in the world) and transaction-time (when Helix learned it); invalidation is append-only.
Importance	A write-time rating (1–10 / 0–1) of how significant a memory is; an input to salience.
Reflection	Periodic synthesis of higher-level insights from clusters of episodes, stored as new linked memories.
Sleep-time consolidation	A background, idle-time worker that consolidates/rewrites memory using a stronger model off the hot path.
Hybrid retrieval	Combining dense (embedding) and sparse (BM25) search.
RRF	Reciprocal Rank Fusion (k=60) — rank-only fusion of multiple result lists.
PPR	Personalized PageRank — lightweight graph-expansion retrieval (HippoRAG-style).
MMR	Maximal Marginal Relevance — diversity/dedup during result selection.
Matryoshka (MRL)	Embeddings whose dimensions can be truncated without re-embedding.
Quantization (int8/binary)	Compressing embedding vectors for smaller/faster storage (with optional float rescore).
Redaction	Tiered removal of secrets/PII (regex → detect-secrets/gitleaks → Presidio) before storage and before any LLM call.
Memory poisoning	An attack that plants false/persistent instructions into long-term memory; guarded by provenance + validation + review.
CRDT	Conflict-free Replicated Data Type — used for mechanical convergence during merge.
3-way semantic merge	Fact/field-level merge against a common ancestor for contradictory facts.
Daemon	The local Helix process (Streamable HTTP on 127.0.0.1) that serves the CLI, MCP shim, and dashboard.
Lethal trifecta	Private data + untrusted content + exfiltration ability — the agent-security risk a memory server must mitigate.