Project-scoped, cross-session episodic memory for AI coding agents (Google Antigravity, Claude Code, Cursor MCP) - local SQLite hybrid keyword+semantic store.
engrim provides 4 well-defined tools with complete input schemas, clear descriptions (194 - 280 chars, within baseline), and documented output schemas. All parameters have types and descriptions. Tool names follow verb_noun convention (engrim_recall, engrim_add, engrim_review). However, descriptions lack explicit WHEN/WHY guidance for LLM selection, and error handling is not documented in tool specs. Output schemas are present but lack field-level descriptions. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.
Write operation: Save a durable memory record to local SQLite storage so it persists across sessions and agent restarts. Call this whenever a non-trivial architectural choice, durable fact, user constraint, or state milestone is established. Appends a new record to the project memory database.
Read-only query: Return this project's curated session-boot memory pack (high-signal architectural decisions, active constraints, and conventions) to orient an agent at the start of a session or after clearing context. Safe to call repeatedly with zero mutations or external side effects. Records are prioritized and truncated to strictly fit within the specified character budget.
Read-only query: Search this project's engrim memory for records relevant to a query using hybrid BM25 keyword and semantic vector ranking. Use before starting non-trivial work to recall prior architectural decisions, technical facts, user feedback, and project state. Safe to invoke repeatedly with zero mutations or side effects.
Read-only check: Inspect conversation and flight recorder history before clearing context to surface recent architectural decisions not yet saved to curated memory. Safe to call with no side effects. Returns heuristic verdict (safe_to_clear: true/false/null) and uncaptured candidate records.
Tool annotations missing: engrim_recall and engrim_context are read-only; engrim_add is destructive. Spec 2026-07-28 supports readOnlyHint and destructiveHint in tool metadata to guide agent planning. None are declared.
Output schema fields lack descriptions. engrim_recall returns 'records' (array of objects) but does not document what fields each record contains (type, summary, detail, tags, timestamp, etc.). LLMs cannot reason about downstream usage without field-level docs.
Error handling not documented in tool specs. No guidance on what errors engrim_add or engrim_recall can return, how to recover, or whether failures are retryable. Agents cannot plan error recovery.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2025-06-18+ | v2 |
Parameter 'origin_agent' in engrim_add uses enum but lacks guidance on when to set it. Should clarify: is this auto-detected from the calling client, or must the agent explicitly pass it? Current description does not explain.
engrim_review description says 'heuristic verdict (safe_to_clear: true/false/null)' but output schema does not document the structure of the response or what 'uncaptured candidate records' looks like. Incomplete output contract.