Structured AI memory layer - single source of truth for what your agent knows :)
Server has 8 tools with complete input schemas and descriptions. Naming follows verb_noun convention (context_*). Descriptions are detailed (100-300 chars) and explain WHEN to use each tool. However, output schemas are not formally documented in the code, responses are inferred from return statements. Parameter descriptions are present but some lack constraint details (e.g., category enum values are documented in description text, not as JSON Schema enum). Error handling is minimal, no recovery guidance or actionable error messages. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Missing pagination for list-like tools. Overall solid foundation but lacks production-grade polish.
Delete a specific entry from a category. Args: category: One of: project, decisions, errors, tasks, ephemeral key: The entry key to delete
Load all entries from a category. Args: category: One of: project, decisions, errors, tasks, ephemeral include_expired: If True, include expired entries (default False)
Log whether recall fired before the first edit in this session. Usually called automatically at the first save - use this tool only when you want the metric for a session that made no edits.
Remove all expired entries across all categories. Call this periodically to keep context lean and token-efficient.
Save or update a context entry. Args: category: One of: project, decisions, errors, tasks, ephemeral key: Short unique identifier e.g. "auth-strategy" or "bug-login-redirect" value: The content to remember. Put the key fact on the first line. tags: Optional list of tags for search e.g. ["auth", "security"] ttl_days: Override default TTL. None = use category default. what_worked: (decisions/errors only) What worked about this entry. what_failed: (decisions/errors only) What failed about this entry. superseded_by: Key of the entry that supersedes this one. The old entry stays but recall will follow this pointer to the successor.
Output schemas not formally documented. Return types are inferred from code (dicts with unspecified fields). LLMs cannot plan downstream calls or validate response structure.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. context_delete and context_purge_expired are destructive but unmarked. Agents cannot reason about side effects.
Category parameter uses free-form string with enum values documented only in description text, not as JSON Schema enum constraint. LLMs may hallucinate invalid categories.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Search across all categories by keyword. Matches against entry keys, values, and tags. Superseded entries are excluded (their successors are returned instead). Args: query: Search term (case-insensitive)
Start a session by loading memory. Call this FIRST, before any task. Returns: - Auto-loaded project and decisions entries (full content) - Compact index of errors, tasks, ephemeral (key + lead fact only) - Recent session recall stats (did recall fire before edits in past sessions?)
Get a summary of all context categories - entry counts, staleness, descriptions. Secondary to context_session_start; use this for quick status checks.
No error handling guidance. Invalid category returns a dict with 'error' key, but no recovery hint (e.g., 'Try one of: project, decisions, errors, tasks, ephemeral'). Agents cannot self-correct.
context_load and context_search return potentially large result sets with no pagination support. No limit parameter, no cursor, no total count. Large categories could exhaust context window.