Neuroscience-grounded persistent memory for AI agents with real emotions, personality, and simulated brain. Exposes MCP tools for remembering and recalling memories with cognitive re-ranking and hybrid retrieval.
Celiums Memory presents a sophisticated neuroscience-grounded memory system with 16 tools, but exhibits significant definition quality gaps. While tool descriptions are generally present (194-char average, within baseline), they lack actionable constraint details, parameter descriptions are often absent or minimal, and output schemas are not documented. Tools are well-named with clear verb-noun patterns (remember, recall, create_claim, etc.), but many parameters lack type definitions visible in the provided schema excerpts. Error handling is minimally documented. The STDIO-only transport and lack of tool annotations further constrain the score. Per-tool analysis reveals inconsistent schema completeness: some tools (remember, recall) have reasonable parameter definitions, but others (checkpoint_export, mcp_call_tool) expose cryptographic keys or raw binary data as parameters, violating security patterns. Most critically, the dual MCP-meta-protocol tools (mcp_tools_list, mcp_call_tool) are inferred rather than directly registered, triggering the inference cap.
Export complete memory state as an encrypted checkpoint artifact with BLAKE3 integrity proof and optional journal replay.
Import a checkpoint artifact to restore memory state. Validates BLAKE3 signatures and decrypts with provided key.
Replay pending journal operations from checkpoint. Used for deterministic recovery after failures.
Query the state of a semantic claim at a specific temporal point with uncertainty quantification.
Create a semantic claim (subject-predicate-object triple) with temporal validity bounds and evidence provenance.
Logically delete a memory. Requires elevated role and confirmation token for audit trail.
Order a list of event IDs by event timestamp and ingestion time for deterministic causality tracking.
Cryptographic keys exposed as tool parameters (checkpoint_export key, checkpoint_import key). Keys should never appear in tool parameters, credentials leak into agent logs and traces.
Output schemas not documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what does recall() return? Is it an array of memory objects with id/content/timestamp fields?). Without response schemas, agents must guess the output structure.
mcp_tools_list and mcp_call_tool are meta-protocol handlers (not business logic tools). Exposing these as MCP tools violates the composition principle, they add noise to the tool list and confuse agents about what the server actually does. These should be implicit MCP protocol mechanics, not selectable tools.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | <=2025-11-25 | v2 |
Retrieve a specific memory by ID with full metadata and cognitive scores.
List all memories in a given scope with optional filtering by tags.
MCP protocol handler: execute a tool by name with JSON arguments.
MCP protocol handler: list available tools with their schemas.
Parse English and Spanish natural language time expressions (e.g., 'yesterday', 'hace 2 semanas') into temporal intervals with uncertainty bounds.
Query memories using hybrid retrieval (semantic + full-text) with cognitive re-ranking based on importance, affect (PAD), and Ebbinghaus retention curves.
Store a memory event with content, tags, and metadata. Supports idempotency keys for deterministic replay and scope-based isolation.
Compute the semantic difference between two claim snapshots, reporting value changes and evidence shifts but filtering out cognitive metadata changes.
Patch memory metadata (tags, importance scores) without modifying content. Content mutations are tracked as separate claim events.
checkpoint_export and checkpoint_import accept raw binary checkpoint_artifact or output_path strings without format validation. No guidance on file formats, size limits, or corruption recovery. Error handling does not explain what to do if BLAKE3 signature validation fails or decryption key is wrong.
Parameter descriptions are absent or minimal for most tools. Examples: event_sequence lacks description for 'event_ids' parameter; semantic_claim_diff lacks description for 'before' and 'after' parameters (what shape are these objects?); delete_memory lacks confirmation details or error guidance for insufficient permissions.
No pagination or result limiting documented for recall() despite description mentioning 'limit' parameter. What happens if limit=10000? Are results streamed or buffered? No guidance on context window impact when returning large memory sets.
delete_memory requires 'elevated role and confirmation token' per description, but parameters show no role or token fields. Either the description is aspirational (not implemented) or parameters are incomplete. Either way, this is a definition gap that breaks the tool's usability.
Tool descriptions mention advanced cognitive features (Ebbinghaus retention curves, PAD affect scoring, importance-weighted re-ranking) but do not document how agents control or interpret these scores. How does an LLM use 'importance' in recall results? Is it 0-1 or 0-100? Can it be filtered?
parse_relative_time accepts 'basis' enum (Utc, Local) but does not document timezone handling or how Local is resolved when running in a distributed system. Ambiguity invites LLM errors in time-critical operations.
create_claim and claim_snapshot use temporal parameters (valid_from_ms, valid_at_ms, known_at_ms) in epoch milliseconds. LLMs frequently miscalculate timestamp conversions. Should return ISO 8601 strings and accept human-friendly time expressions via parse_relative_time integration.