Skip to content
ExoVault

Search Documentation

Search across all documentation pages

Glossary

Definitions of key terms used throughout ExoVault documentation.

A#

Agent Key#

An API authentication token with the exv_ prefix. Agent keys are scoped to specific integrations and vaults, with configurable permission levels (read_only, read_write, admin). The raw key is never stored -- only its SHA-256 hash.

Agent Integration#

A configuration that represents an AI agent connected to ExoVault. Each integration has an agent type, label, scope preset, vault access rules, and a wrapped Master Encryption Key (MEK). Multiple agent keys can be generated per integration.

Agent Session#

A tracked period of agent activity. Sessions are created automatically when agents make API calls with an x-agent-run-id header. Sessions track tool calls, duration, and can be viewed in the dashboard.

Archive#

Soft-deletion of a memory. Archived memories are excluded from search results by default (unless includeArchived: true is set) but remain in the database and can be restored.

B#

Blind Index#

A privacy-preserving search technique. Content is tokenized into words/n-grams, each token is HMAC-signed with the MEK, and the resulting hashes are stored. Queries are tokenized the same way, enabling keyword-level matching without decrypting content.

Broadcast Message#

A message sent to all agents by using targetId: "*". Broadcast messages remain in "pending" status even after delivery, so multiple agents can pick them up.

C#

Confidence#

A score from 1-5 indicating how certain the agent is about a memory's accuracy. 1 = speculative, 5 = certain. Set based on whether the information was directly stated, inferred, or guessed.

Constraint#

A memory type representing a mandatory rule or restriction. Constraints are loaded with high priority during session_start (up to 10 per profile) because violating them is considered an error.

Content Hash#

A SHA-256 hash of memory content + type, used for exact-match deduplication without comparing encrypted text. Catches identical duplicates at zero API cost.

Context Checkpoint#

A bulk operation that saves multiple memories and a session summary at once. Used at session end to persist everything the agent learned. Supports 4-layer deduplication and idempotency.

Context Profile#

A preset that controls how much context is loaded during session_start. Profiles include default, planning, incident, handoff, deep, minimal, and none.

Correction#

A memory type that records when the user corrected the agent. High importance (4-5) to ensure the agent remembers what was wrong and what is right.

D#

Daily Digest#

An automated daily summary of all agent sessions from the previous day, generated at 3 AM UTC as a single episodic memory. Consolidates multiple sessions into one overview so agents have a concise "yesterday" context.

Deduplication (Dedup)#

A multi-layer system that prevents duplicate memories. Layers: (1) in-batch content hash, (2) database content hash, (3) blind token overlap, (4) semantic embedding similarity. Enabled by setting dedup: true.

Document (Vault Document)#

Configurable text documents stored per-vault: soul (agent identity), instructions (operating rules), skills (capabilities), checks (validation rules). Documents support inheritance from global settings and can be agent-editable.

E#

Embedding#

A 3,072-dimensional vector representation generated by Google's gemini-embedding-2-preview model. Natively multimodal — text, images, audio, video, and PDFs are embedded in the same vector space. Used for semantic similarity search and cross-modal retrieval. Stored in pgvector columns.

Entity#

A named item extracted from memory content (e.g., technology names, project names, people). Stored as an array on the memory for filtering and search improvement.

Episodic Memory#

A memory type representing a session summary or event record. Created automatically by context_checkpoint from the sessionSummary field. Provides "what happened last time" context.

Extraction Pipeline#

A background processing system that analyzes conversation turns for extractable information. Uses signal detection to identify high-value turns, then processes them to create memories.

F#

Fact#

A memory type representing objective, verifiable information. The most common memory type, typically about the project, environment, or user.

H#

Hooks#

Shell scripts that fire on agent lifecycle events (SessionStart, PromptSubmit, Stop for Claude Code; beforeSubmitPrompt, afterAgentResponse for Cursor). Used for context injection (loading relevant memories into prompts) and turn capture (recording conversations for extraction). Installed per-project by exovault connect.

I#

Idempotency#

The ability to safely retry an operation without creating duplicates. Achieved via externalWriteId on write-memory and create-task, or auto-generated checkpoint:{agentRunId}:{index} keys on context-checkpoint.

Importance#

A score from 1-5 indicating how critical a memory is. 1 = trivial, 5 = must always remember. Higher-importance memories are prioritized in search results and context loading.

Inngest#

A background job processing platform used by ExoVault for asynchronous tasks: embedding generation, extraction pipeline processing, and webhook delivery.

K#

Knowledge Graph#

A graph structure connecting memories and notes with typed, directional relations. Enables multi-hop traversal for contextual discovery. Nodes are memories and notes; edges are typed links.

An edge in the knowledge graph connecting two nodes (memories or notes) with a typed relation. Types include wiki_link, derived_from, contradicts, refines, part_of, supersedes, source_of, references, and manual.

M#

Media#

Video, audio, image, or PDF files attached to memories or notes. Encrypted at rest in Supabase Storage. Automatically processed: embedded via Gemini for cross-modal search, text extracted via Gemini 2.5 Flash (speech transcription, visual descriptions, OCR).

Master Encryption Key (MEK)#

A per-integration 256-bit key used for AES-256-GCM encryption of content. The MEK is wrapped (encrypted) with the Server Encryption Key and stored on the integration.

MCP (Model Context Protocol)#

A protocol for connecting AI models with external tools and data sources. ExoVault provides an MCP server (exovault-mcp-server) that exposes all functionality as MCP tools.

Memory#

A unit of knowledge stored in ExoVault. Each memory has a type (fact, skill, preference, constraint, task, episodic, correction), encrypted content, importance, confidence, and optional metadata.

Memory Unit#

A composite usage metric that combines embedding tokens and operations for quota tracking.

MMR (Maximal Marginal Relevance)#

A re-ranking algorithm that reduces redundancy in search results by penalizing items similar to already-selected items. Controlled by the diversity parameter.

Multimodal#

The ability to process and search across different content types (text, video, audio, images, PDFs) in a unified vector space. ExoVault uses Gemini's natively multimodal embeddings so a text query can find relevant video content and vice versa.

N#

Note#

An encrypted document stored in ExoVault with title, content, and tags. Notes can be organized in folders and linked to memories via the knowledge graph.

P#

Preference#

A memory type representing a user's subjective choice or opinion. Preferences should influence agent behavior and are loaded with high priority.

R#

RLS (Row Level Security)#

PostgreSQL-level access control that ensures users can only access their own data. All tables with user data have RLS policies. The service role key bypasses RLS for server-side operations.

RRF (Reciprocal Rank Fusion)#

A score fusion algorithm that combines rankings from multiple search signals (vector, blind index, graph) into a single unified ranking. Each signal contributes based on its weight.

S#

Scope#

A permission level on an agent key: read, write, search, or admin. Scopes can be configured globally per integration and overridden per vault.

Server Encryption Key (SEK)#

The top-level encryption key stored as the ENCRYPTION_KEY environment variable. Used to wrap/unwrap per-integration MEKs. Loss of this key makes all encrypted data unrecoverable.

Session Summary#

A text summary of a conversation session, provided in context_checkpoint. Stored as an episodic memory with an auto-generated headline for display.

Signal Detection#

A zero-LLM-cost analysis of conversation turns that identifies high-value content for the extraction pipeline. Uses keyword matching and pattern detection to assign a signal score.

Significance Gate#

A filter that prevents low-value sessions from generating episodic memories. Only sessions with meaningful activity (non-trivial tool usage, memory writes, or substantial conversation) pass the gate and produce episodic records. Reduces memory noise.

Skill#

A memory type representing learned patterns, techniques, or project-specific knowledge. Skills help the agent work more effectively in future sessions.

Supersession#

The process of replacing an outdated memory with a newer one. The old memory is archived and linked to the new one. Can be triggered explicitly or automatically via deduplication.

T#

Task#

A memory type with additional status tracking (backlog, todo, in_progress, done, blocked). Tasks can be assigned to specific agents and have completion criteria.

Temporal Decay#

A search ranking adjustment that reduces the score of older memories relative to newer ones. Controlled by decayHalfLife (default: 30 days).

V#

Vault#

An encrypted container for organizing memories, notes, and documents. Each user can have multiple vaults. Agent access can be restricted to specific vaults.

Vault Document#

See Document.

W#

A [[double-bracket]] reference in content that is automatically parsed and converted to a wiki_link knowledge graph relation.

Wrapped MEK#

The Master Encryption Key encrypted (wrapped) with the Server Encryption Key. Stored as wrappedMek + wrappedMekIv on the agent integration. Unwrapped at request time for content encryption/decryption.