A signed, append-only record of the decisions behind AI-agent work — the decision, the reasoning, and who wrote it down, in the repository where the work happens.
mnema has 25 tools with consistent naming (verb_noun pattern) and descriptions present for all. However, critical gaps undermine quality: NO input schemas are visible in the provided source code, only parameter names and types are mentioned in comments. Parameter descriptions are absent from the schema definitions. Output schemas are completely undocumented. Error handling and recovery guidance are not evident. The codebase shows domain sophistication (state machines, audit trails, proof chains) but the MCP surface lacks the rigor required for production agent use. Average tool score: 42/100.
Get accountability across all trees (who did what, auditor's view).
Get antipatterns across all trees (events with a given shape, auditor's view).
Get exposure across all trees separately (auditor's view, keeps trees apart).
Get the reference graph across all trees (auditor's view).
Get the timeline of all events across all trees (auditor's view).
Get the opening context: where the actor left off, what work is live, which patterns are adopted, which decisions govern, and what is waiting on somebody to rule on it.
Capture a memory (an observation without a decision context).
NO input schemas visible in source code. All 25 tools lack formal JSON Schema definitions with type constraints, required fields, and parameter descriptions. Only informal parameter lists in comments.
Output schemas completely undocumented. No tool describes what fields it returns, their types, or structure. LLMs cannot plan downstream tool calls or extract required data.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | <=2025-11-25 | v2 |
Create a skill (a reusable pattern or capability).
Create a task (a unit of work with a state machine).
Move a decision to a new state (via the state machine gate).
Narrow the opening context to one actor's view.
Get the rules that govern a path in the record.
Dry-run a move: simulate a transition and return the verdict without writing.
Link knowledge (create a cross-reference between records).
Get the next actions available for a task (what moves the gate allows).
Read one whole record by id (the full body and metadata).
Record a decision (an ADR-style decision record with reasoning and status).
Record a handoff (a knowledge fact about work passed to another agent).
Record an observation (a knowledge fact about the world).
Get the context an actor left off in (their last run).
Get the rules in force at a path and record the asking (read that also writes).
Search records across all visible trees (index of what matched).
Move a skill to a new state (via the state machine gate).
Serve a pattern's body and record that it was served (the one read that also writes).
Move a task to a new state (via the state machine gate).
Parameter descriptions missing from schema. Tools like 'search' accept 'limit' (default 10, max 100) but this constraint is not documented in the parameter description, only in the informal spec above.
No error handling or recovery guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next (e.g., 'Record not found. Try search() first.').
Ambiguous tool names for state transitions. 'task_transition', 'decision_transition', 'skill_transition' are similar but operate on different entity types. LLMs may conflate them. Consider 'transition_task_state', 'transition_decision_state', 'transition_skill_state' for clarity.