Loading ToolBench
Unified persistent memory system using SQLite with sqlite-vec and FTS5. Provides semantic search, vector embeddings, and archival capabilities for long-term context retention.
Ember3 has 15 tools with complete input schemas and descriptions, but quality is inconsistent. Tool names follow verb_noun convention (ember_learn, archive_search, deep_recall) which is good. However, descriptions are often generic and lack actionable context for LLM selection. Parameters have types and descriptions, but many lack constraints (enums, ranges, format specs). Output schemas are not documented, critical for agent chaining. Error handling guidance is absent. The server exposes internal mechanics (vectors, cells, centroids) in descriptions, violating the instruction to hide implementation details. No tool supports dry-run or confirmation for destructive operations (ember_delete). Parameter relationships (e.g., top_k defaults, project scoping) are underdocumented.
Regex or fuzzy pattern search across archived content
Read archive chunks with start/end line navigation for large content
Keyword search across archived content using FTS5
Save session logs, debug records, and reference documents to the archive
Search both memory and archive simultaneously using semantic similarity
Automatically learn from user's first message with semantic understanding
Correct or supersede old information with new facts
No output schemas documented. LLMs cannot infer what fields to expect from tool responses, breaking agent chaining and forcing discovery detours.
Destructive tool (ember_delete) lacks confirmation/dry-run support. Agents can permanently delete memories without verification, risking data loss.
Parameter constraints missing. 'top_k' has no min/max bounds; 'importance' and 'chunk_type' lack enum definitions. LLMs will hallucinate invalid values.
Descriptions expose internal mechanics (vectors, cells, centroids, HESTIA scoring). Per instructions, these should be hidden from LLMs. Descriptions should focus on user intent, not implementation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Delete a memory permanently
Traverse knowledge graph edges between related memories
Store facts, preferences, decisions, or corrections as persistent memories
List memories with optional filtering by tags or importance
Retrieve the full content of a specific memory by ID
Search persistent memory using semantic similarity
Recompute vector space centroids for clustering
Get statistics about the memory system
No error handling guidance. Tools lack recovery hints (e.g., 'If memory not found, try ember_recall with a broader query'). Agents cannot self-correct on failures.