Local-first AI long-term memory with FSRS scheduling, git-backed snapshots, and semantic diffing for efficient context management.
18 tools with complete schemas and descriptions, but significant gaps in parameter documentation and error handling. Tool names follow verb_noun convention (remember, recall, search, forget, etc.), which is good. However, many parameter descriptions are minimal or missing entirely. For example, 'remember' lacks descriptions for 'difficulty', 'tags', 'overwrite', and 'confirm' parameters. Output schemas are not documented, callers don't know what fields to expect. Error handling is absent: no guidance on retryability, user-fixable vs fatal errors, or recovery paths. The 'summarize' tool requires ANTHROPIC_API_KEY but doesn't document this dependency or error cases. Pagination is implemented (list, recall_all) but inconsistently, some tools lack limit/page params. No tool declares permissions or security implications. Baseline: 100% of A+ tools have param descriptions; this server has ~40% coverage.
Show semantic diff between current memories.json and a commit hash.
List recent git-backed snapshots of memories.
Preview or restore memories.json from a backup commit hash.
Bulk forget memories by ids (array).
Bulk apply review rating to multiple memories.
Export all memories for a project as a markdown file in the data directory.
Delete an outdated memory.
Parameter descriptions missing or incomplete. 'difficulty', 'tags', 'overwrite', 'confirm' in 'remember' lack descriptions. 'force' in 'bulk_forget' and 'forget' undescribed. Baseline: 100% of A+ tools have param descriptions.
Output schemas not documented. Callers don't know what fields 'remember', 'recall', 'list', 'stats', 'export' return. This forces LLMs to guess field names for downstream tool calls.
No error handling guidance. Tools don't document what errors are retryable, user-fixable, or fatal. 'summarize' requires ANTHROPIC_API_KEY but doesn't document failure modes if key is missing.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
List all memories for a project with health scores.
Permanently purge trashed memories older than 30 days.
Get fading memories for a project (below 85% health).
Recall fading memories across all projects (optional tag).
Save a decision to long-term memory. ALWAYS include the project name.
Rename a project (change project field on all memories).
Restore a soft-deleted memory by id.
Update memory health after recall.
Search memories by keyword in a project.
Return project/global stats (health, counts, reviews).
Summarize all memories for a project into a 3-sentence brief using Anthropic (requires ANTHROPIC_API_KEY).
Destructive operations lack confirmation pattern. 'forget', 'bulk_forget', 'purge' can permanently delete data but don't support dry-run or confirmation step. Agents can accidentally destroy memories.
No permission declarations or security scopes. Tools don't declare what permissions they require (e.g., 'write:memory', 'delete:memory'). Audit trail and least-privilege configuration impossible.