Long-term, drift-free cognitive memory engine for AI Agents and Cursor IDE.
MCP Server: io.github.novotapeter.bs/cortex
The MCP server io.github.novotapeter.bs/cortex provides a “long-term, drift-free cognitive memory engine” intended for use with AI agents and the Cursor IDE. Its purpose is to support persistent memory for these environments, as described in its repository metadata.
🛠️ Key Features
Long-term cognitive memory
Drift-free behavior
Built for AI Agents and Cursor IDE
🚀 Use Cases
Supplying durable memory for AI agents
Enabling cognitive memory within Cursor IDE workflows
⚡ Developer Benefits
Targeted for long-term memory needs in agent/IDE integrations
Explicit focus on reducing drift over time
⚠️ Limitations
Available metadata does not list tools, endpoints, or implementation details.
No additional operational constraints are provided in the source data.
Most memory systems treat everything as fluid vector soup—causing agents to hallucinate, forget strict technical requirements, or confuse user preferences with code specs.
Blue Synergy Cortex implements an enterprise-grade Tri-Stream Architecture backed by transactional PostgreSQL and enterprise knowledge graph architecture:
User Stream: Generalizes conversational habits and user preferences over time.
Identity Stream: Deterministic guardrails and behavioral axioms that prevent sycophancy.
Knowledge Stream (Living Specs): Exact architecture specifications, hardware limits, and system parameters organized into an authoritative Universal Domain Ontology. This stream is protected against lossy compression, semantic decay, and accidental drift.
Quickstart (1 Minute Setup)
API Key Required: Blue Synergy Cortex is a fully hosted, zero-maintenance cognitive cloud engine. You need an active API key to connect your client.
Cortex is not a passive key-value store; it is an active cognitive operating system with strict governance boundaries. Understanding when and why to call each tool ensures high-fidelity memory consolidation.
The Standard Agent Turn Cycle
For every user turn, your agent should inherently follow this golden loop:
At turn start: Call cortex_recall to fetch ground-truth context (User Bio + Active Project Specs + Graph Relations).
Generate response: Formulate the answer grounded in the retrieved memory.
At turn end: Call cortex_ingest_turn to send the turn to the asynchronous ingestion buffer.
Detailed Tool Reference
Tool Name
Type
When to Call
System Behavior
cortex_recall
query (string) compact (bool, optional)
Before answering the user (mandatory for full context awareness).
Sub-65ms hybrid vector and topological graph search. Injects verified user traits, active project specifications, and cross-entity graph facts. Set compact: true to strip profile preambles and return only high-relevance facts (recommended for CLI / Claude Code).
cortex_ingest_turn
Ingestion
Immediately after answering every conversational turn. Crucial for agent behavioral distillation and user biography generation.
Asynchronously ingests the full conversational turn. The Memory Firewall filters dialogue noise, learns communication preferences, and stages technical facts for scheduled consolidation.
cortex_create_project
Provisioning
When the user declares, initializes, or approves a new project scope.
Explicitly registers and authorizes a new canonical project_id. Creates the project baseline, prevents future memory writes from falling back into general, and automatically adopts related staged notes from the general pool.
cortex_store_memory
Write
For explicit system constraints, business rules, and technical decisions that must be guaranteed and stored in the KNOWLEDGE stream.
Curated persistent write. Safety Gate: If project_id is not yet authorized by the user, facts are safely staged in the general memory pool (general) to prevent project pollution.
cortex_get_project_spec
Inspection
When you need the complete Living Canonical Specification of a project.
Deterministically retrieves the verified, unified specification (Living Document) of the project. Note: Returns only authorized and consolidated specifications—not unorganized scratchpad notes.
cortex_manage_draft
Governance
When a new product or domain is proposed and awaits explicit user authorization.
Approves quarantined project drafts (draft_*) into active production specifications, or permanently purges rejected concepts. Promotes staged facts from general into a dedicated project scope.
cortex_triage_general
Maintenance
When the unstructured staging pool (general) accumulates multi-project notes.
Semantic clustering engine that organizes notes in general, pairs them with corresponding projects, and executes workspace compaction.
Project Governance & Lifecycle
To prevent memory fragmentation and unauthorized project sprawl, Cortex treats memory with strict governance boundaries:
text
[Incoming Note / Message]
│
▼
Is Project Authorized?
├── NO ──► Staged in General Pool ('general')
│ │
│ ▼
│ Awaits Triage via `cortex_triage_general`
│ (or explicitly authorized via `cortex_create_project`)
│ │
└── YES ◄─────────┘
│
▼
[Asynchronous Consolidation Cycle]
│
▼
[Active Living Specification & Knowledge Topology]
Why is my project_spec empty after storing notes?
Cortex enforces a Human-in-the-Loop policy. If a project_id has not yet been explicitly authorized by the user, incoming facts are safely staged in the general memory pool to prevent polluting active workspaces.
Once the project is approved (via cortex_manage_draft or explicit creation), the nocturnal consolidation engine automatically synthesizes these staged notes into the official Living Specification (cortex_get_project_spec).
Core Architecture Highlights
Sub-65ms Zero-LLM Recall: Direct high-performance vector and topological graph matching—completely bypassing slow, non-deterministic runtime LLM keyword extractors.
Dual-Zone Fast-Append: Instant raw fact capture (< 20ms) into an inbox staging zone, followed by scheduled asynchronous consolidation into an immutable, unified Living Specification.
Semantic Invariant Guard: Algorithmic verification gate that enforces factual density and continuity, mathematically preventing the truncation or loss of critical parameters during re-indexing.
Immunity Shield: Operational errors, tool crashes, transient hallucinations, and conversational apologies are filtered out by design and never committed to long-term memory.
Support & Feedback
Encountered an issue or have a feature suggestion?
