MCP server exposing YantrikDB cognitive memory engine as tools and resources for Claude Desktop, Hermes Agent, and other MCP clients
YantrikDB MCP server demonstrates solid tool design with comprehensive schemas and descriptions across 23 tools. All tools have descriptions (avg ~150 chars) and input schemas with typed parameters. Naming follows verb_noun convention (memory_record, entity_relate, claim_attach). However, several tools lack output schema documentation, some descriptions are generic, and error handling guidance is minimal. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Parameter descriptions are present but could be more prescriptive about constraints and formats.
Attach structured claims (subject-relation-object triples) to a memory for knowledge graph grounding and contradiction detection.
Get all claims attached to a memory.
Search for claims by subject, relation, or object pattern.
Get database statistics: total memories, namespaces, entities, claims, and storage size.
Optimize the database by reclaiming unused space. This may take time on large databases.
Get an entity and its relationships from the knowledge graph.
Create or update a relationship between two entities (people, places, concepts, etc.) in the knowledge graph.
Search for entities by name pattern or semantic similarity.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications. Tools marked DESTRUCTIVE (memory_delete, namespace_delete, observer_reset) should declare destructiveHint=true for LLM safety.
Output schemas not documented in tool definitions. LLMs cannot plan downstream calls or extract required fields (e.g., what does memory_record return? A rid? Metadata?). Responses must include structured field definitions.
Error handling lacks recovery guidance. No indication of what LLM should do on failure (retry? ask user? call alternative tool?). E.g., memory_get with invalid rid should suggest memory_search or memory_list as alternatives.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Delete a memory by its ID. This is irreversible.
Get a specific memory by its ID.
List memories with optional filtering by namespace, domain, source, or memory type. Returns paginated results.
Search memories by semantic similarity to a natural language query. Uses multi-signal scoring: vector similarity, temporal decay, recency, importance, and optional knowledge graph expansion.
Refine a previous recall with a follow-up query. Use this when recall returned low-confidence results and hints suggested rephrasing. Combines the original query embedding with the refinement embedding (weighted: 0.4 original + 0.6 refinement) and excludes already-seen memory IDs.
Store a new memory in the cognitive memory engine.
Update an existing memory's metadata (importance, valence, domain, etc.). The text itself is immutable; to change content, record a new memory and optionally deprecate the old one.
Create a new namespace for memory isolation.
Delete a namespace and all its memories. This is irreversible.
List all namespaces in the database.
Compute derived signals from recent observations (user engagement, preference signals, anomalies, etc.).
Record a system event for passive observation and autonomous learning.
Get recent system events with optional filtering by event kind, time range, or app.
Reset the observer (clear all events and counters). This is irreversible.
Get a summary of observed events: total counts, rates, circadian distribution, and derived signals.
Parameter descriptions lack prescriptive constraints. E.g., 'importance' accepts 0.0-1.0 but description doesn't state range. 'memory_type' accepts 'episodic|semantic|procedural' but no enum constraint visible. LLMs need formal constraints, not prose hints.
Pagination support incomplete. memory_list and claim_search accept limit/offset but no total_count or next_cursor returned. Large result sets risk context window exhaustion without clear pagination semantics.