Persistent memory framework for LLM agents: episodic and semantic memory backed by Postgres + pgvector, exposed as an MCP server for Claude Code and other assistants
mnemos has 5 well-named tools with clear, action-oriented names (remember, recall, list, forget, reflect). All tools have descriptions (avg 180 chars, within baseline 194). Input schemas are present with typed parameters and descriptions. However, output schemas are NOT documented, tools return free-text strings rather than structured objects, forcing LLMs to parse unstructured responses. Error handling is minimal; tools lack recovery guidance. Parameter descriptions are adequate but lack format constraints (e.g., top_k range, fact_id format). No tool annotations (readOnlyHint, destructiveHint) despite clear risk profiles.
Forget (archive) a specific memory by its ID, as shown by mnemos_list_memories. Soft-delete: the fact is archived, not erased, and stays visible in mnemos's audit log — never call this speculatively, only when the user explicitly asks to forget something.
List all currently active memories with their IDs, for browsing or as input to mnemos_forget.
Retrieve memories relevant to a query — call this near the start of a task that might benefit from prior context about the user or their preferences/projects.
Run a consolidation pass: merge near-duplicate memories, decay stale ones, archive anything that decays past the forget threshold. Requires a real LLM provider configured (ANTHROPIC_API_KEY or GROQ_API_KEY) for the merge step. Call this occasionally to keep memory tidy, not on every turn.
Store a durable fact worth recalling in a future session — a stated preference, a decision and its reasoning, a recurring convention. Not routine task details, not anything already in the codebase/CLAUDE.md.
Output schemas not documented. All tools return free-text strings (e.g., 'Remembered [UUID]: fact') instead of structured JSON objects. LLMs cannot parse responses reliably or chain downstream calls.
No tool annotations. Tools have clear risk profiles (mnemos_remember=WRITE, mnemos_forget=REVERSIBLE, mnemos_reflect=WRITE) but lack readOnlyHint/destructiveHint/idempotentHint annotations in schema.
Minimal error handling. Tools lack recovery guidance. E.g., mnemos_reflect returns 'LLM_PROVIDER is set to mock' but does not suggest next steps (set ANTHROPIC_API_KEY or GROQ_API_KEY). No actionable error messages.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Parameter constraints missing. top_k defaults to 5 but no min/max bounds documented. fact_id is a string but format (UUID) not stated in description. limit defaults to 50 but no upper bound specified.
No pagination support in list_memories. Returns up to 50 items by default but no next_cursor or total_count. Large memory stores will be truncated without visibility into remaining items.