mindgraph

A structured semantic memory graph for agentic systems, built in Rust with CozoDB as the embedded Datalog storage engine.

Overview

mindgraph provides a typed, versioned knowledge graph organized into six conceptual layers:

Layer	Purpose	Example Node Types
Reality	Raw observations & sources	Source, Snippet, Entity, Observation
Epistemic	Reasoning & knowledge	Claim, Evidence, Hypothesis, Theory, Concept
Intent	Goals & decisions	Goal, Project, Decision, Option, Constraint
Action	Affordances & workflows	Affordance, Flow, FlowStep, Control
Memory	Persistence & recall	Session, Trace, Summary, Preference
Agent	Control plane	Agent, Task, Plan, Approval, Policy

The graph supports 48 built-in node types and 70 built-in edge types, each with type-safe property structs. User-defined custom types are also supported via the CustomNodeType trait.

Features

Type-safe schema -- 48 node types and 70 edge types as Rust enums with typed props, plus extensible Custom(String) variants
CozoDB storage -- Embedded Datalog database with SQLite persistence or in-memory mode
Full-text search -- FTS indices on node labels and summaries with scoring and type/layer filters
Structured filtering -- NodeFilter builder for type (single or multi-type), layer, label substring, prop value, and confidence range queries
Graph traversal -- Optimized 2-query BFS, reasoning chains, neighborhoods, path finding, subgraph extraction, weight threshold filtering
Builder pattern -- Ergonomic fluent API for node and edge updates
Pagination -- Bounded result sets with has_more detection for production use
Batch operations -- Multi-row inserts (chunked at 100) and GraphOp-based batch apply
Versioning -- Append-only version history for both nodes and edges, with point-in-time snapshots
Tombstone cascade -- Soft-delete a node and all connected edges in one call
Data lifecycle -- purge_tombstoned() for hard-deleting old data; export()/import() for graph snapshots; backup()/restore_backup() for file-level backups
Provenance tracking -- Link extracted knowledge to its sources
Entity resolution -- Alias table, fuzzy matching, merge_entities() for deduplication
Multi-agent support -- AgentHandle provides scoped per-agent identity; all mutations auto-set changed_by, with sub_agent() for hierarchical agents
Custom node/edge types -- CustomNodeType trait for compile-time registration of user-defined types with typed ser/de
Default agent identity -- set_default_agent() reduces boilerplate in builder patterns
Confidence & salience -- Validated 0.0-1.0 scores on all nodes and edges
Thread safety -- MindGraph is Send + Sync, safe to share via Arc<MindGraph> or into_shared()
Async support -- Optional AsyncMindGraph wrapper for tokio runtimes (feature flag: async) with all methods
Server-side query filtering -- Query patterns push filtering into CozoDB Datalog for efficient large-graph queries
Embedding/vector search -- Pluggable EmbeddingProvider (sync) and AsyncEmbeddingProvider (native async) traits, CozoDB HNSW indices, semantic_search() with cosine distance
Salience decay -- Exponential decay with configurable half-life via decay_salience(), plus auto_tombstone() for cleanup
Event subscriptions -- on_change() callbacks, on_change_filtered() with EventFilter, and watch() async streaming via broadcast channels
Convenience constructors -- add_claim(), add_entity(), add_goal(), add_observation(), add_session(), add_preference(), add_summary(), add_link()
Graph statistics -- stats() returns comprehensive GraphStats with counts by type/layer
Enhanced query composition -- OR filters, time ranges, salience ranges, prop conditions, graph-aware connected_to filter
Typed export/import -- export_typed() / import_typed() with TypedSnapshot for structured graph transfer
Validated batch -- validate_batch() pre-validates operations before apply_validated_batch()
OpenAI embeddings -- Optional openai feature flag for OpenAIEmbeddings provider via ureq
Tracing integration -- Optional tracing feature flag for observability instrumentation on key graph methods
Production-safe async -- AsyncMindGraph returns Error::TaskJoin instead of panicking on spawn failures

Quick Start

use mindgraph::*;

fn main() -> Result<()> {
    // Open a persistent graph (SQLite-backed)
    let graph = MindGraph::open("my_graph.db")?;
    // Or in-memory for testing:
    // let graph = MindGraph::open_in_memory()?;

    // Add a claim node
    let claim = graph.add_node(
        CreateNode::new("Rust is memory safe", NodeProps::Claim(ClaimProps {
            content: "Rust is memory safe".into(),
            claim_type: Some("factual".into()),
            ..Default::default()
        }))
        .confidence(Confidence::new(0.95)?)
    )?;

    // Add supporting evidence
    let evidence = graph.add_node(
        CreateNode::new("Borrow checker", NodeProps::Evidence(EvidenceProps {
            description: "Borrow checker prevents dangling pointers".into(),
            ..Default::default()
        }))
    )?;

    // Connect with a typed edge (evidence supports claim)
    graph.add_edge(CreateEdge::new(
        evidence.uid.clone(),
        claim.uid.clone(),
        EdgeProps::Supports { strength: Some(0.9), support_type: Some("empirical".into()) },
    ))?;

    // Update using the builder pattern
    graph.update(&claim.uid)
        .confidence(Confidence::new(0.99)?)
        .changed_by("agent-1")
        .reason("strong supporting evidence")
        .apply()?;

    // Traverse the reasoning chain (includes start node at depth 0)
    let chain = graph.reasoning_chain(&claim.uid, 5)?;
    assert_eq!(chain[0].node_uid, claim.uid); // start node
    assert_eq!(chain[0].depth, 0);

    Ok(())
}

Async Usage

Enable the async feature for tokio integration:

[dependencies]
mindgraph = { version = "0.6", features = ["async"] }

use mindgraph::*;

#[tokio::main]
async fn main() -> Result<()> {
    let graph = AsyncMindGraph::open_in_memory().await?;

    let node = graph.add_node(
        CreateNode::new("Async claim", NodeProps::Claim(ClaimProps {
            content: "Works in async contexts".into(),
            ..Default::default()
        }))
    ).await?;

    // AsyncMindGraph is Clone (wraps Arc<MindGraph>),
    // so it can be shared across tasks
    let g = graph.clone();
    let handle = tokio::spawn(async move {
        g.count_nodes(NodeType::Claim).await
    });
    assert_eq!(handle.await.unwrap()?, 1);

    // For updates, use update_node/update_edge directly
    // (builder types hold references and can't cross await points)
    graph.update_node(
        node.uid,
        Some("Updated claim".into()),
        None, None, None, None,
        "agent".into(), "async update".into(),
    ).await?;

    Ok(())
}

Custom Types

Define your own node types without forking the crate:

use mindgraph::*;
use serde::{Serialize, Deserialize};

#[derive(Debug, Clone, Serialize, Deserialize)]
struct CodeSnippet {
    language: String,
    source: String,
}

impl CustomNodeType for CodeSnippet {
    fn type_name() -> &'static str { "CodeSnippet" }
    fn layer() -> Layer { Layer::Reality }
}

let graph = MindGraph::open_in_memory().unwrap();
let node = graph.add_custom_node("hello.rs", CodeSnippet {
    language: "rust".into(),
    source: "fn main() {}".into(),
}).unwrap();

// Type-safe deserialization
let snippet: CodeSnippet = node.custom_props().unwrap();
assert_eq!(snippet.language, "rust");

Breaking change in v0.6: NodeType and EdgeType no longer implement Copy (they implement Clone). Add .clone() where needed.

Multi-Agent Support

Use AgentHandle to scope operations to a specific agent identity:

use std::sync::Arc;
use mindgraph::*;

let graph = Arc::new(MindGraph::open_in_memory().unwrap());
let alice = graph.agent("alice");

// All mutations auto-set changed_by to "alice"
let node = alice.add_entity("My Entity", "test").unwrap();
let my_nodes = alice.my_nodes().unwrap();
assert_eq!(my_nodes.len(), 1);

// Sub-agents for hierarchical systems
let sub = alice.sub_agent("alice-summarizer");
assert_eq!(sub.parent_agent(), Some("alice"));

Event Streaming

Filter and stream graph events (requires async feature):

use mindgraph::*;

let graph = MindGraph::open_in_memory().unwrap();

// Sync filtered callback
let filter = EventFilter::new().event_kinds(vec![EventKind::NodeAdded]);
graph.on_change_filtered(filter, |event| {
    println!("New node: {}", event);
});

With async streaming:

// AsyncMindGraph::watch() returns a WatchStream
let stream = async_graph.watch(
    EventFilter::new()
        .event_kinds(vec![EventKind::NodeAdded])
        .layers(vec![Layer::Epistemic])
);
// stream.recv().await returns filtered events

Tracing

Enable the tracing feature for observability:

[dependencies]
mindgraph = { version = "0.6", features = ["tracing"] }

Key methods (add_node, search, find_nodes, reachable, stats, etc.) are instrumented with tracing::instrument. Combine with tracing-subscriber to get structured logs.

API Reference

MindGraph

The main entry point. All operations go through this struct. It is Send + Sync and can be shared across threads via Arc<MindGraph>.

Construction:

Method	Description
`MindGraph::open(path)`	Open a persistent SQLite-backed graph
`MindGraph::open_in_memory()`	Create an in-memory graph (for testing)
`into_shared()`	Wrap in `Arc<MindGraph>` for sharing across threads
`set_default_agent(name)`	Set default agent identity for builder fallbacks
`default_agent()`	Get current default agent identity
`storage()`	Access the underlying `CozoStorage` for advanced Datalog queries
`agent(name)`	Create a scoped `AgentHandle` (requires `Arc<MindGraph>`)
`nodes_by_agent(agent_id)`	Get all live nodes created by a specific agent

Convenience constructors:

Method	Description
`add_claim(label, content, confidence)`	Add a Claim node with defaults
`add_entity(label, entity_type)`	Add an Entity node with defaults
`add_goal(label, priority)`	Add a Goal node with defaults
`add_observation(label, description)`	Add an Observation node with defaults
`add_session(label, focus)`	Add a Session node with defaults
`add_preference(label, key, value)`	Add a Preference node with defaults
`add_summary(label, content)`	Add a Summary node with defaults
`add_memory(label, content)`	Deprecated -- use `add_session()` instead
`add_link(from, to, edge_type)`	Add an edge with default props for the edge type
`add_custom_node::<T>(label, props)`	Add a node with a user-defined custom type

Node operations:

Method	Description
`add_node(CreateNode)`	Add a new node (auto-assigns UID, version 1)
`add_nodes_batch(Vec<CreateNode>)`	Bulk insert multiple nodes (multi-row, chunked at 100)
`get_node(uid)`	Get a node by UID, returns `None` if not found
`get_live_node(uid)`	Get a node, errors if not found or tombstoned
`update_node(uid, ...)`	Update fields directly (increments version)
`update(uid)`	Begin a builder-pattern update, finalize with `.apply()`
`node_exists(uid)`	Check if a live node exists (O(1), no deserialization)
`count_nodes(node_type)`	Count live nodes of a given type
`count_nodes_in_layer(layer)`	Count live nodes in a given layer

Edge operations:

Method	Description
`add_edge(CreateEdge)`	Add a new edge (validates both endpoints are live)
`add_edges_batch(Vec<CreateEdge>)`	Bulk insert edges (validates all endpoints first)
`get_edge(uid)`	Get an edge by UID, returns `None` if not found
`get_live_edge(uid)`	Get an edge, errors if not found or tombstoned
`update_edge(uid, ...)`	Update fields directly (increments version)
`update_edge_builder(uid)`	Begin a builder-pattern update, finalize with `.apply()`
`edges_from(uid, edge_type?)`	Get all live edges from a node, optionally filtered by type
`edges_to(uid, edge_type?)`	Get all live edges to a node, optionally filtered by type
`count_edges(edge_type)`	Count live edges of a given type
`get_edge_between(from, to, edge_type?)`	Find edges between two nodes, optionally by type

Traversal:

Method	Description
`reachable(uid, opts)`	BFS to find all nodes reachable through filtered edge types
`reasoning_chain(uid, max_depth)`	Traverse epistemic edges; returns start node at depth 0
`neighborhood(uid, depth)`	Get all nodes within `depth` hops in any direction
`find_path(from, to, opts)`	Find the actual shortest path between two nodes
`subgraph(uid, opts)`	Extract all reachable nodes and their interconnecting edges

Tombstone operations:

Method	Description
`tombstone(uid, reason, by)`	Soft-delete a node with audit trail
`restore(uid)`	Restore a tombstoned node
`tombstone_edge(uid, reason, by)`	Soft-delete an edge with audit trail
`restore_edge(uid)`	Restore a tombstoned edge
`tombstone_cascade(uid, reason, by)`	Tombstone a node and all connected edges

Version history:

Method	Description
`node_history(uid)`	Get full version history (create, updates, tombstone)
`edge_history(uid)`	Get full version history for an edge
`node_at_version(uid, version)`	Get the JSON snapshot at a specific version number

Search & filtering:

Method	Description
`search(query, opts)`	Full-text search across labels/summaries with FTS scoring
`find_nodes(filter)`	Structured filtering by type, layer, label, props, confidence
`find_nodes_paginated(filter)`	Same as above with `Page<GraphNode>` pagination metadata

Data lifecycle:

Method	Description
`purge_tombstoned(older_than)`	Hard-delete tombstoned data (and associated versions/aliases/provenance)
`export()`	Export entire graph as a `GraphSnapshot`
`import(snapshot)`	Import a graph snapshot (additive merge)
`backup(path)`	Backup database to a file
`restore_backup(path)`	Restore database from a backup file

Provenance & entity resolution:

Method	Description
`add_provenance(record)`	Link a node to its extraction source
`add_alias(text, canonical_uid, score)`	Register an alias for entity resolution
`resolve_alias(text)`	Resolve text to a canonical entity UID
`aliases_for(uid)`	List all aliases for a canonical entity, sorted by score
`merge_entities(keep, merge, reason, by)`	Merge two entities: retarget edges/aliases, tombstone duplicate
`fuzzy_resolve(text, limit)`	Substring match on alias text

Embedding/vector search:

Method	Description
`configure_embeddings(dimension)`	Initialize HNSW index for semantic search
`embedding_dimension()`	Get configured embedding dimension (None if not configured)
`set_embedding(uid, vec)`	Store an embedding vector for a node
`get_embedding(uid)`	Retrieve a node's embedding vector
`delete_embedding(uid)`	Remove a node's embedding
`semantic_search(query_vec, k)`	Find k nearest neighbors by cosine distance (auto-compensates for tombstoned nodes)
`embed_node(uid, provider)`	Generate and store embedding via `EmbeddingProvider`
`embed_nodes(uids, provider)`	Bulk embed multiple nodes via `embed_batch()`, skips tombstoned
`semantic_search_text(query, k, provider)`	Embed query text and search

Salience decay:

Method	Description
`decay_salience(half_life_secs)`	Apply exponential decay to all live nodes
`auto_tombstone(min_salience, min_age_secs)`	Tombstone old nodes below salience threshold

Event subscriptions:

Method	Description
`on_change(callback)`	Subscribe to graph mutation events, returns `SubscriptionId`
`on_change_filtered(filter, callback)`	Subscribe with `EventFilter` for selective events
`watch(filter)`	(async feature) Create a `WatchStream` for async event streaming
`unsubscribe(id)`	Remove a subscription

Statistics:

Method	Description
`stats()`	Get comprehensive `GraphStats` (counts by type, layer, embeddings, etc.)

Utility:

Method	Description
`list_nodes(pagination)`	List all live nodes with pagination
`clear()`	Delete all data from all relations (for testing/reset)

Typed export/import:

Method	Description
`export_typed()`	Export live graph as `TypedSnapshot` with structured nodes/edges/embeddings
`import_typed(snapshot)`	Import a typed snapshot (additive merge, skips existing UIDs, restores embeddings)

Batch operations (GraphOp):

Method	Description
`batch_apply(ops)`	Execute a batch of AddNode/AddEdge/Tombstone operations
`validate_batch(ops)`	Pre-validate a batch (auto-assigns UIDs, tracks cross-refs), returns `ValidatedBatch`
`apply_validated_batch(batch)`	Apply a pre-validated batch

Query patterns (server-side filtered via CozoDB Datalog):

Method	Description
`active_goals()`	Goals with `status == "active"`, ranked by priority
`pending_approvals()`	Approvals with `status == "pending"`, sorted by requested_at
`unresolved_contradictions()`	CONTRADICTS edges with `resolution_status == "unresolved"`
`open_decisions()`	Decisions with status `"open"` or `"deliberating"`
`open_questions()`	OpenQuestions with status `"open"` or `"partially_addressed"`
`weak_claims(threshold)`	Claims with `confidence < threshold`, sorted ascending
`nodes_in_layer(layer)`	All live nodes in a given layer

Paginated variants:

Method	Description
`nodes_in_layer_paginated(layer, page)`	Paginated nodes in a layer
`edges_from_paginated(uid, edge_type?, page)`	Paginated edges from a node
`edges_to_paginated(uid, edge_type?, page)`	Paginated edges to a node
`weak_claims_paginated(threshold, page)`	Paginated weak claims
`active_goals_paginated(page)`	Paginated active goals, sorted by priority in DB

AsyncMindGraph

Available behind the async feature flag. Wraps Arc<MindGraph> and exposes async versions of all methods via tokio::task::spawn_blocking.

Method	Description
`AsyncMindGraph::open(path)`	Async open
`AsyncMindGraph::open_in_memory()`	Async in-memory open
`AsyncMindGraph::from_sync(graph)`	Wrap an existing `MindGraph`
`inner()`	Access the underlying `&MindGraph`

AsyncMindGraph is Clone and can be shared across tokio tasks. All methods from MindGraph are available as async variants, taking owned arguments instead of references.

Note: The builder types (NodeUpdate, EdgeUpdate) hold references and cannot cross .await points. Use update_node() / update_edge() directly in async code.

Builders

CreateNode -- built with CreateNode::new(label, props), with optional chained methods:

.summary(text) -- set the node summary
.confidence(Confidence) -- set epistemic certainty (default 1.0)
.salience(Salience) -- set contextual relevance (default 0.5)
.privacy(PrivacyLevel) -- set privacy level (default Private)
.with_uid(Uid) -- pre-assign a UID (for cross-referencing in validate_batch)

CreateEdge -- built with CreateEdge::new(from_uid, to_uid, props), with optional chained methods:

.confidence(Confidence) -- set edge confidence (default 1.0)
.weight(f64) -- set edge weight (default 0.5)

NodeUpdate -- started with graph.update(uid):

graph.update(&uid)
    .label("Updated label")
    .summary("New summary")
    .confidence(Confidence::new(0.9)?)
    .salience(Salience::new(0.8)?)
    .changed_by("agent-1")
    .reason("new evidence")
    .apply()?;

EdgeUpdate -- started with graph.update_edge_builder(uid):

graph.update_edge_builder(&edge_uid)
    .weight(0.95)
    .confidence(Confidence::new(0.9)?)
    .changed_by("agent-2")
    .reason("re-evaluated")
    .apply()?;

Traversal

Control traversal behavior with TraversalOptions:

use mindgraph::*;

let opts = TraversalOptions {
    direction: Direction::Both,         // Outgoing, Incoming, or Both
    edge_types: Some(vec![              // None = follow all edge types
        EdgeType::Supports,
        EdgeType::Refutes,
    ]),
    max_depth: 5,                       // BFS depth limit
    weight_threshold: Some(0.5),        // None = no weight filter
};

let steps = graph.reachable(&start_uid, &opts)?;
for step in &steps {
    // node_type is NodeType enum, edge_type is Option<EdgeType>
    println!("depth {}: {} ({:?}) via {:?}, parent: {:?}",
        step.depth, step.label, step.node_type, step.edge_type, step.parent_uid);
}

PathStep includes parent_uid for backtracking. find_path uses this to return only the nodes on the actual shortest path (not all reachable nodes).

Pagination

Use Pagination for bounded result sets:

use mindgraph::*;

// First page of 10 items
let page1 = graph.nodes_in_layer_paginated(Layer::Epistemic, Pagination::first(10))?;
assert!(page1.items.len() <= 10);

// Next page
if page1.has_more {
    let page2 = graph.nodes_in_layer_paginated(
        Layer::Epistemic,
        Pagination { limit: 10, offset: 10 },
    )?;
}

Core Types

Type	Description
`Uid`	UUID v4 identifier for nodes and edges (inner field is private)
`Confidence`	Validated f64 in 0.0-1.0 (epistemic certainty)
`Salience`	Validated f64 in 0.0-1.0 (contextual relevance, decays over time)
`PrivacyLevel`	`Private`, `Shared`, or `Public`
`Timestamp`	Unix timestamp as f64
`NodeProps`	Discriminated union of all 48 node property structs
`EdgeProps`	Discriminated union of all 70 edge property structs

Schema

48 node types across 6 layers:

Layer	Node Types
Reality (4)	Source, Snippet, Entity, Observation
Epistemic (24)	Claim, Evidence, Warrant, Argument, Hypothesis, Theory, Paradigm, Anomaly, Method, Experiment, Concept, Assumption, Question, OpenQuestion, Analogy, Pattern, Mechanism, Model, ModelEvaluation, InferenceChain, SensitivityAnalysis, ReasoningStrategy, Theorem, Equation
Intent (6)	Goal, Project, Decision, Option, Constraint, Milestone
Action (5)	Affordance, Flow, FlowStep, Control, RiskAssessment
Memory (5)	Session, Trace, Summary, Preference, MemoryPolicy
Agent (8)	Agent, Task, Plan, PlanStep, Approval, Policy, Execution, SafetyBudget

70 edge types across categories:

Category	Edge Types
Structural (5)	ExtractedFrom, PartOf, HasPart, InstanceOf, Contains
Epistemic (31)	Supports, Refutes, Justifies, HasPremise, HasConclusion, HasWarrant, Rebuts, Assumes, Tests, Produces, UsesMethod, Addresses, Generates, Extends, Supersedes, Contradicts, AnomalousTo, AnalogousTo, Instantiates, TransfersTo, Evaluates, Outperforms, FailsOn, HasChainStep, PropagatesUncertaintyTo, SensitiveTo, RobustAcross, Describes, DerivedFrom, ReliesOn, ProvenBy
Provenance (5)	ProposedBy, AuthoredBy, CitedBy, BelievedBy, ConsensusIn
Intent (9)	DecomposesInto, MotivatedBy, HasOption, DecidedOn, ConstrainedBy, Blocks, Informs, RelevantTo, DependsOn
Action (5)	AvailableOn, ComposedOf, StepUses, RiskAssessedBy, Controls
Memory (5)	CapturedIn, TraceEntry, Summarizes, Recalls, GovernedBy
Agent (10)	AssignedTo, PlannedBy, HasStep, Targets, RequiresApproval, ExecutedBy, ExecutionOf, ProducesNode, GovernedByPolicy, BudgetFor

Architecture

mindgraph
├── graph.rs          -- MindGraph: the main public API + NodeUpdate/EdgeUpdate builders
├── async_graph.rs    -- AsyncMindGraph: tokio wrapper (behind "async" feature)
├── storage/
│   ├── cozo.rs       -- CozoStorage: CozoDB CRUD, traversal, pagination, batch ops
│   └── migrations.rs -- Schema DDL (CozoDB :create statements + indices)
├── schema/
│   ├── mod.rs        -- Layer, NodeType (48), EdgeType (70) enums
│   ├── node.rs       -- GraphNode, CreateNode
│   ├── edge.rs       -- GraphEdge, CreateEdge
│   ├── node_props.rs -- NodeProps discriminated union
│   ├── edge_props.rs -- EdgeProps discriminated union
│   └── props/        -- Per-layer property structs
│       ├── reality.rs    (4 structs)
│       ├── epistemic.rs  (24 structs)
│       ├── intent.rs     (6 structs)
│       ├── action.rs     (5 structs)
│       ├── memory.rs     (5 structs)
│       └── agent.rs      (8 structs)
├── traversal.rs      -- Direction, TraversalOptions, PathStep
├── query.rs          -- Pagination, Page<T>, GraphStats, DecayResult, TypedSnapshot, etc.
├── types.rs          -- Uid, Confidence, Salience, PrivacyLevel, Timestamp
├── provenance.rs     -- ProvenanceRecord, ExtractionMethod
├── embeddings.rs     -- EmbeddingProvider (sync) + AsyncEmbeddingProvider traits
├── events.rs         -- GraphEvent, EventKind, EventFilter, SubscriptionId
├── watch.rs          -- WatchStream (async filtered event stream, behind "async")
├── agent.rs          -- AgentHandle (scoped per-agent graph access)
├── openai.rs         -- OpenAIEmbeddings (behind "openai" feature)
└── error.rs          -- Error types + Result alias

Storage

CozoDB is used as the embedded storage engine. It runs Datalog queries over relations stored in SQLite (persistent) or in-memory (testing). The schema defines six core relations:

Relation	Purpose	Key
`node`	All graph nodes with universal metadata	`uid`
`edge`	All graph edges with typed properties	`uid`
`node_version`	Append-only node version snapshots	`(node_uid, version)`
`edge_version`	Append-only edge version snapshots	`(edge_uid, version)`
`provenance`	Extraction lineage records	`(node_uid, source_uid)`
`alias`	Entity resolution mappings	`(alias_text, canonical_uid)`
`mg_meta`	Key-value config store (e.g., embedding dimension)	`key`
`node_embedding`	Vector embeddings with HNSW index (created on demand)	`uid`

Indices are created for edge traversal (from_uid, to_uid), node lookup (node_type, layer), provenance queries, and alias resolution.

Design Decisions

Props as JSON columns -- Node and edge properties are stored as JSON in CozoDB, with NodeProps/EdgeProps Rust enums providing type safety at the API boundary. This allows CozoDB Datalog to filter on props fields using get(props, 'field', default) without schema migration.
Tombstoning over deletion -- Soft-delete preserves audit trails. Tombstoned entities are excluded from live queries but remain accessible for forensic review. tombstone_cascade removes a node and all its edges atomically.
Append-only versioning -- Every mutation to a node or edge creates a new version snapshot, enabling full history reconstruction and point-in-time queries via node_at_version.
2-query BFS traversal -- Graph traversal fetches all live edges in one query, runs BFS in-memory, then batch-fetches node metadata in a second query. This reduces traversal from O(N) database queries to exactly 2, regardless of graph size. Recursive CozoDB Datalog was tested but found unreliable across versions.
Server-side filtering -- Query patterns like active_goals() and weak_claims() push filtering into CozoDB Datalog rather than loading all nodes into memory. Paginated variants (e.g., active_goals_paginated) sort in the database before applying :limit/:offset.
Tombstone sentinel -- tombstone_at uses 0.0 as the sentinel value for "not tombstoned" since CozoDB columns use fixed types. All live-query filters check tombstone_at == 0.0.
Thread safety -- MindGraph is Send + Sync. CozoDB's DbInstance uses internal locking, so Arc<MindGraph> works safely across threads.
Async via spawn_blocking -- AsyncMindGraph wraps Arc<MindGraph> and delegates all operations to tokio::task::spawn_blocking. This avoids blocking the tokio runtime while leveraging CozoDB's synchronous API.
Private Uid inner field -- Uid(String) keeps its inner field private to prevent accidental construction of invalid UIDs. Use Uid::new(), Uid::from(), or Uid::as_str().

License

MIT

mindgraph 0.6.0