Loading ToolBench
Epistemic Bio-Homeostatic Identity & Memory Kernel for AI Agents with Model Context Protocol (MCP) Interface
Soul Engine MCP provides 21 tools with complete input schemas and descriptions. Naming follows verb_noun convention (soul_*) consistently. However, descriptions are often verbose (200-400 chars) and lack actionable guidance for LLM selection. Parameter descriptions exist but many lack constraints (enums, ranges, formats). Output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; most tools return generic error dicts without recovery guidance. Security-critical tools (soul_memory_delete, soul_review_commit) require human_event_ref authorization but lack clear permission documentation. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels.
Check the status and heartbeat counters of the background SoulDaemon supervisor.
Get a comprehensive memory digest of active reviewed memories and current identity state, optionally filtered by plan_id or agent_id.
Run an offline dream consolidation cycle to prune decayed memory traces and simulate scenarios. Optionally provide scenario_prompt, outcomes, and task_context.
Score pending imagined outcomes against a realized external result. Compares dream predictions with actual outcomes.
Retrieve the current Soul's identity state, including soul_version, constitution_version, traits, neuromodulators (dopamine, cortisol, serotonin), narrative, unresolved_tensions, state_hash, and created_at.
Run the self-healing routine to verify and repair any state hash inconsistencies.
Output schemas not documented. LLMs cannot predict response structure for any tool. soul_get_identity, soul_recall, soul_digest, soul_review_status return complex nested objects (neuromodulators dict, memories array, digest object) but no schema is provided.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools marked as READ_ONLY, WRITE, REVERSIBLE, DESTRUCTIVE in comments but not exposed to MCP client. soul_memory_delete and soul_review_commit are destructive but lack annotation.
Parameter constraints missing. soul_reward accepts valence and confidence as floats but no min/max bounds documented. soul_recall limit defaults to 5 but no range specified. soul_review_stage_decision decision enum not enforced in schema.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | <=2025-11-25 | v2 |
Log a host event (human, agent, tool, environment, or system origin) in a tamper-evident log with SHA-256 hash chain.
Execute a GDPR salted privacy erasure on a memory. Permanently revokes cryptographic recoverability using salt nullification and content redaction.
Rollback the active reviewed memory set to a target version using human authorization. Creates a new monotonic version V+1 with audit receipts.
Search and retrieve memories using semantic or hybrid search. Returns matching memories with optional filtering by plan_id or agent_id.
Trigger metacognitive reflection on unresolved tensions in the Soul's memory graph. Optionally provide interpretations to guide reflection.
Ingest an experience or fact into the Soul's episodic memory (quarantine). Accepts content, source_kind, provenance, entity_key, and other metadata.
Atomically commit a review cycle with human confirmation. Requires commit_human_event_ref for authorization.
Generate the atomic preview diff and SHA-256 preview hash for a review cycle before committing.
Stage a human review decision on a candidate memory. Decision can be 'remember', 'correct', 'reject', or 'replace_old'. Requires human_event_ref for authorization.
Start a Soul Review Cycle for a session. Trigger kind can be 'explicit', 'timeout', or 'daemon'.
Get the current status and candidate list for a review cycle in a session.
Process a reward signal to update the Soul's neuromodulators and traits. Accepts source, valence (-1.0 to 1.0), confidence (0.0 to 1.0), and optional task_context, evidence_receipt, review_receipt.
Rollback the Soul's identity state to a target version without erasing audit history.
Update a Soul trait to a new value. Requires trait name and new_value. Agents cannot self-attest human approval.
Verify an experience or claim against the Soul's knowledge base using epistemic verification.
Error handling lacks recovery guidance. Tools return generic {'error': str(exc)} dicts. No categorization (retryable vs fatal), no actionable next steps. soul_verify, soul_reflect, soul_heal provide no guidance on failure modes.
Permission documentation incomplete. soul_memory_delete and soul_review_commit require human_event_ref but no scope/permission declaration (e.g., 'requires: delete:memory', 'requires: review:commit'). Unclear what permissions agents need.