Personal Knowledge Augmentation System for AI coding agents - MCP server implementation providing document search, indexing, knowledge management, tool intelligence tracking, and memory management capabilities
Engram defines 10 tools with action-based parameters rather than verb-prefixed names. All tools have descriptions (194-char baseline met), but schemas are action-dispatch patterns that obscure individual operation contracts. Parameters lack type constraints (enums for actions, ranges for limits). No output schemas documented. Error handling and recovery guidance absent. Tool names (docs, entities, sessions, tools, memory, obligations, search, knowledge, coordination, work) are nouns, not verbs, violates pattern:tool naming convention. The action parameter pattern (e.g., action: 'search'|'plan'|'index') is generic and forces LLMs to reason about string values rather than discrete tool boundaries.
Manage session coordination and handoffs (Layer 5). Coordinate between sessions, manage handoffs, and track coordination state.
Search, index, and manage documents in the knowledge base. Supports planning document ingestion, indexing documents from paths, searching indexed documents, and managing document lifecycle.
Manage knowledge graph entities (Layer 1). Create, retrieve, update, and delete entities representing concepts, tools, people, and other domain objects. Support entity relationships and aliases.
Manage document intelligence and knowledge extraction (Layer 6). Extract knowledge from documents, manage document metadata, and track document relationships.
Manage Memory OS - persistent memory system for agents (Layer 8). Create, retrieve, update, and delete memory items. Support memory hierarchies, evidence tracking, and memory status management.
Manage agent obligations and commitments. Create, track, and resolve agent obligations with status management and resolution tracking.
Tool names are nouns (docs, entities, sessions) instead of verb_noun pattern (search_docs, create_entity, list_sessions). LLMs cannot infer action intent from name alone and must parse the action parameter string.
Action parameters accept free-form strings (e.g., action: 'search'|'plan'|'index') without enum constraints. LLMs may hallucinate invalid actions like 'analyze' or 'summarize'. No validation or error guidance provided.
No output schemas documented for any tool. LLMs cannot predict response structure, forcing them to guess field names for downstream tool chaining. Violates pattern:tool and pattern:response-shaper.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 45 | <=2025-11-25 | v2 |
Unified cross-layer search across all knowledge layers. Search entities, documents, sessions, memories, and other indexed content with semantic and keyword search capabilities.
Manage session history and context (Layer 2). Create, retrieve, and manage sessions that track conversation history, context, and state across interactions.
Track tool usage, learn preferences, and get recommendations (Layer 4). Log tool usage outcomes, track tool switches, get recommendations based on context, and view tool statistics.
Manage projects and tasks (Layer 7). Create, retrieve, and manage work items, projects, and task hierarchies.
Numeric parameters (limit) lack min/max constraints. No guidance on pagination, result limits, or handling large datasets. Agents may request 1M results, causing timeouts or context overflow.
No error handling or recovery guidance. Tools do not document what errors are retryable, user-fixable, or fatal. No actionable error messages shown in schema.