v0.2.0

0.2.0 - 2026-02-11

Added

• getEntityEdge, deleteEntityEdge, and clearGraph methods on GraphitiClient — wrapping the previously unused get_entity_edge, delete_entity_edge, and clear_graph MCP server tools
• entity_types filter parameter on searchNodes — filters results by entity classification (e.g. Preference, Organization, Procedure)
• center_node_uuid parameter on searchFacts — anchors fact search around a specific entity node ("tell me everything about X")
• entity_types and center_node_uuid optional parameters on the memory_recall agent tool, threaded through to Graphiti search calls
• memory_forget now supports deleting individual facts via a new fact_id parameter — fetches the fact to determine its group_id, checks canWriteToGroup authorization, then deletes the entity edge
• graphiti-mem fact <uuid> CLI command — fetch and display a specific fact (entity edge) by UUID as JSON
• graphiti-mem clear-graph CLI command — clear graph data for specified groups; requires --confirm safety flag, supports --group <id...> to target specific groups
• graphiti-mem cleanup CLI command to find and optionally delete orphaned Graphiti episodes — episodes that exist in Graphiti but have no source_group relationship in SpiceDB (can happen when Phase 2 of a two-phase import fails). Supports --group, --last, --delete, and --dry-run flags.
• readRelationships method on SpiceDbClient for querying existing authorization tuples by resource type, relation, and subject filter
• Bulk import: graphiti-mem import now uses two-phase approach — Graphiti ingestion first, then a single BulkImportRelationships streaming RPC to SpiceDB (with batched WriteRelationships fallback), replacing per-file interleaved writes
• ZedToken consistency tuning: SpiceDB reads after writes now use at_least_as_fresh consistency with the token from the preceding write, ensuring causal consistency without the cost of fully_consistent; reads without a prior write use minimize_latency for optimal performance

Changed

• memory_forget tool: episode_id is now optional — callers must provide either episode_id or fact_id (not both)
• Docker image updated from stale ghcr.io/getzep/graphiti-mcp:latest to zepai/knowledge-graph-mcp:latest, matching the upstream Graphiti project's published image

Fixed

• UUID mismatch between Graphiti and SpiceDB: addEpisode now polls getEpisodes in the background to resolve the real server-side UUID assigned by Graphiti (the MCP add_memory tool only returns a "queued" message, not the UUID). SpiceDB authorization relationships are written with the real UUID once resolved, fixing broken fragment-level authorization (memory_forget, canDeleteFragment, lookupViewableFragments). The polling is non-blocking — memory_store and auto-capture return immediately while SpiceDB writes happen in the background; graphiti-mem import awaits all UUIDs before Phase 2 bulk write.
• memory_forget now uses filter-based DeleteRelationships RPC to clean up SpiceDB relationships, fixing orphaned tuples when deleting fragments stored to non-default groups or by other subjects
• Flaky "batch conversation capture extracts entities" E2E test — replaced fixed 20s sleep with a polling loop (up to 8 attempts at 5s intervals) to handle variable Graphiti extraction queue latency

Internal

• Extracted mcpUrl getter in GraphitiClient for single-point trailing slash adjustment across all 4 MCP endpoint references