An MCP server that gives any MCP client (Claude Code, Claude Desktop, Cursor) anatid memory. Embedded bitemporal graph memory for AI agents, built on DuckDB.
anatid MCP server has well-structured tool definitions with complete schemas and descriptions for all 13 tools. Naming follows verb_noun conventions (remember, recall, relate, forget, etc.). Descriptions are substantive (avg ~150 chars) and explain WHAT the tool does and WHEN to use it. All parameters have type definitions and descriptions. However, several tools lack actionable error guidance, and some descriptions could better explain prerequisites or dependencies. Output schemas are not explicitly documented in the source. Risk annotations (WRITE, READ_ONLY, DESTRUCTIVE) are present but not formally mapped to MCP's destructiveHint/readOnlyHint. No dry-run or confirmation patterns for irreversible operations (forget, prune) despite their destructive nature.
Commit a proposed patch from ingest, unchanged or edited, in one transaction
What is known about one entity
Supersede a memory and close and open its edges, in one transaction
Stop believing a memory (soft) or erase it (hard). With hard=True, deletes the row, its edges, its embedding and its provenance.
A note in, a reviewed patch of facts and edges applied as one. Proposes a patch from text using a configured model.
Where a memory came from: its chain, episodes and writers
Remove memories and edges older than a cutoff date, optionally by kind or entity
Destructive operations (forget, prune) lack dry-run confirmation or explicit user confirmation step. Agents can permanently erase data without a safety gate.
Risk annotations (WRITE, READ_ONLY, DESTRUCTIVE) are present in metadata but not formally mapped to MCP tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLM clients cannot parse risk from the schema.
Output schemas are not documented in tool definitions. LLMs cannot predict what fields recall(), context(), or provenance() return, forcing them to guess downstream field names.
Error handling descriptions are absent. Tools do not explain what to do if a memory_id is invalid, an entity is not found, or a patch fails to apply.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Hybrid recall (vector + BM25 + graph, fused with RRF). Combines BM25 text search with a two-hop walk of the entity graph, so it returns facts about things connected to the seed even when those facts contain none of the search words.
Connect two entities, so facts about one are reachable from the other
Write one memory, optionally about named entities
Raw SQL escape hatch for read-only queries. Opt-in via ANATID_ENABLE_SQL=1. Runs arbitrary read-only SQL with enforced limits.
Replace a memory with a corrected one, keeping the old version
Retire the connection between two entities
Parameter dependencies are undocumented. E.g., correct() has remove_relations and add_relations arrays with nested objects, but no description of the required structure or what happens if both are empty.