Glossary

A plain-language dictionary of the terms used throughout this wiki. Aimed at readers who are new to semantic memory, knowledge graphs, or AI agent infrastructure.

A

• Abstention — the retrieval engine’s ability to say “I don’t know.” When confidence is too low, /recall returns {decision: "low_confidence", hits: []} instead of a confidently wrong top-1 result.
• Audit chain — an append-only log where each row stores the SHA-256 hash of the previous row, so any modification or deletion is detectable.

B

• Bearer token — a secret string sent in the Authorization header to authenticate a request. Brain Server supports opaque bearer tokens (default) and JWT/JWS.
• Bi-temporal — recording both when a fact is valid in the world (valid_at/invalid_at) and when it was recorded (observed_at). Enables point-in-time recall.
• BM25 — the classic lexical scoring function (term-frequency × inverse-document-frequency) used by SQLite’s FTS5 full-text index.

C

• Capacity envelope — a configurable bound on docs / DB size / RSS. Writes that exceed it return HTTP 507; reads are never blocked.
• Chunk — a unit of memory stored in a knowledge row. Text is split into chunks by a CommonMark-aware splitter (heading-boundary splits, code-fence-safe).
• CommonMark — a standard, unambiguous specification of Markdown. Brain Server’s chunker uses a CommonMark parser so all constructs are handled correctly.
• Connector — a supervised ingester (e.g. GitHub issues) that backfills external sources through the source/revision pipeline.
• CSP (Content Security Policy) — an HTTP header controlling what resources a page may load. Brain Server serves a strict CSP for the API and a relaxed one for the WASM client.

D

• Decision path / trace — the recorded record of a recall: injected chunks, fused scores, abstention decision, access scope, principal, and domains searched. Replayable via GET /recall/{trace_id}/trace.
• Domain — a scoped memory namespace (health, business, code…) with its own knowledge graph. Retrieval auto-routes between domains by centroid and falls back on a miss.
• DSAR — Data Subject Access Request. Brain Server’s /dsar workflow locates → exports → purges → issues a chain-verifiable deletion certificate.

E

• Embedding — a numeric vector representing text, such that semantically similar texts are close in vector space. Brain Server uses static embeddings (model2vec) — no transformer forward pass.
• Egress — data leaving your device/network. Brain Server has no data egress by default.
• Evidence — the verbatim snippet, line span, source link, and highlight ranges attached to a retrieved chunk — what a result is actually based on.

F

• FTS5 — SQLite’s full-text-search index, scored with BM25. The lexical retrieval leg.
• Fusion — merging multiple ranked lists into one. Brain Server uses Reciprocal Rank Fusion.

G

• Graph leg — the optional third retrieval leg: Personalized PageRank over the knowledge graph, opt-in via ?graph=true.
• Governance — the layer that keeps memory honest and auditable: audit log, quarantine, write-back gating, DSAR, retention.

H

• Hybrid retrieval — combining vector (semantic) and lexical (keyword) search. Brain Server runs both legs concurrently and fuses them.
• Hub dampening — a technique that reduces the influence of very-high-degree graph nodes (mega-hubs), so taxonomy tag clouds don’t drown out real semantic edges.

I

• Ingest — the act of adding memory: POST /ingest, /ingest/memory, or /ingest/markdown.

J

• JWT / JWS — JSON Web Token / JSON Web Signature. The opt-in enterprise authentication mode. Only RS256/ES256/EdDSA allowed (never HS256 or none).

K

• Knowledge graph — entities and the relationships between them, extracted from markdown links. Traversable and queryable.
• KNN — k-nearest-neighbors, the vector search that finds the closest embeddings to a query.

L

• LexSpec — the structured lexical query: terms, quoted phrases, exclusions (-"..."), and exact code paths.
• Loopback — 127.0.0.1, the local machine. Brain Server is loopback-safe by default (refuses 0.0.0.0 unless BIND_PUBLIC=1).

M

• MCP — Model Context Protocol, a standard for exposing tools to agents. Brain Server ships an mcp binary.
• Multi-domain — running several scoped domain databases that auto-route and cross-reference on a miss.

P

• PII — personally identifiable information. Brain Server applies deterministic read-time output redaction to PII; there is no write-time placeholder vault (v1.20.19).
• PRF — pseudo-relevance feedback: deterministic query expansion that fires only when the top result appears in both retrieval legs within a bounded rank.
• Proposal — a write-back candidate scored by the server but held in a queue until a human approves it. Nothing enters memory autonomously.
• Provenance — per-retriever ranks, fused score, expansion terms, and evidence attached to each result.

Q

• QueryDoc — the structured query document accepted by /recall (query, filters, provenance flag, graph flag).

R

• Recall — retrieval. POST /recall is the primary endpoint.
• Reciprocal Rank Fusion (RRF) — a deterministic, weight-free merge: score = Σ 1/(k + rank), with k = 60.
• Retention — how long data is kept. Content is kept until purged; audit rows honor BRAIN_AUDIT_RETENTION_DAYS if set.

S

• Span verification — POST /verify checks whether a claim is literally supported by a chunk’s text (deterministic lexical match, no LLM).
• Static embedding model — a model with no transformer forward pass, just token lookup (model2vec / potion-retrieval-32M). Cheap on CPU.
• Supersede — marking a new fact as replacing an old one. Atomically expires the old fact from current recall; historical recall still returns it.
• SQLite vec0 — a SQLite extension for vector search (KNN over quantized embeddings).

T

• Temporal evidence — the observed_at / valid_from / valid_to / authority stamps that make point-in-time recall possible.
• Tombstone — a hash-only record left when data is purged, proving a deletion occurred.
• Trace — see Decision path.

U

• Untrusted-evidence boundary — the OWASP LLM01:2025 pattern where every retrieved result serializes untrusted: true, signaling the consuming agent to treat it as untrusted evidence.

V

• Vector — see Embedding.
• vec0 KNN — the vector search leg over quantized embeddings.

W

• WAL — Write-Ahead Logging, SQLite’s concurrency mode used by Brain Server (with a busy timeout so concurrent writers queue rather than fail).
• Write-back gate — the human-in-the-loop mechanism that scores a candidate but requires approval before it becomes memory.