Friday MCP server

Friday

Self-hosted persistent cognitive memory layer for AI coding agents.

Persists architecture decisions, schemas, and constraints across sessions via the Model Context Protocol (MCP).

---

The Problem: Session Amnesia

Modern AI coding agents (Cursor, Claude Code, Antigravity, VS Code) excel at isolated code generation. However, in continuous engineering workflows, developers encounter a structural limitation: Session Amnesia.

Current workarounds fall into three deeply flawed patterns:

                  ┌─────────────────────────────────────────────────────────┐
                  │          WHY STANDARD APPROACHES BREAK DOWN             │
                  └─────────────────────────────────────────────────────────┘

   1. Context Windows (RAM)           2. Static Rules Files             3. Standard Vector RAG
  ┌─────────────────────────┐       ┌─────────────────────────┐       ┌─────────────────────────┐
  │ • Ephemeral volatile    │       │ • Linear token tax      │       │ • Matches text phrasing,│
  │   memory (clears on     │       │   (2,500 tokens burned  │       │   NOT system topology   │
  │   every new thread)     │       │   on every trivial fix) │       │ • Blind to directed     │
  │ • Lost-in-the-middle    │       │ • Stale rules accumu-   │       │   call graphs & schema  │
  │   degradation on 50k+   │       │   late & conflict       │       │   dependencies          │
  │   token prompts         │       │ • Zero cross-tool sync  │       │ • Hallucinates blast    │
  │ • High latency & cost   │       │   (Cursor ≠ Claude CLI) │       │   radii of refactors    │
  └─────────────────────────┘       └─────────────────────────┘       └─────────────────────────┘

1. Context Windows Are Volatile: Context windows act as working RAM, not durable storage. Clearing a thread or restarting an agent resets state. Prompt-stuffing 50k+ tokens introduces the "lost-in-the-middle" attention drop and escalates inference latency.
2. Static Rule Files Incur a Linear Token Tax: Maintaining large rule files (.cursorrules, AGENTS.md) forces the model to re-read thousands of lines on every keystroke, leading to contradictory instructions and cross-editor fragmentation.
3. Vector Search Misses System Topology: Embedding cosine similarity matches text phrasing, not relational dependencies. Vector search cannot traverse directed graphs: $$\text{Table: accounts} \longrightarrow \text{FK: subscriptions} \longrightarrow \text{Service: BillingService} \longrightarrow \text{Worker: InvoicePoller}$$

---

Architecture: Multi-Layer Cognitive Substrate

Friday runs as a self-hosted background service providing a structured, four-tier memory substrate accessed via the Model Context Protocol (MCP):

┌────────────────────────────────────────────────────────────────────────────────────────┐
│               AI CODING CLIENTS (Cursor / Claude Code / Antigravity / VS Code)         │
└───────────────────────────────────────────┬────────────────────────────────────────────┘
                                            │
                                7 MCP Tools (stdio / HTTP)
                                ├── add_memory       (persist decisions & rationale)
                                ├── add_fact         (versioned immutable truths)
                                ├── memory_search    (targeted semantic recall)
                                ├── get_context      (compiled multi-layer prompt)
                                ├── get_blast_radius (graph dependency traversal)
                                ├── delete_memory    (evict obsolete episodic memories)
                                └── revoke_fact      (prune conflicting/invalid truths)
                                            │
                                            ▼
┌────────────────────────────────────────────────────────────────────────────────────────┐
│                                 FRIDAY COGNITIVE ENGINE                                │
│                                                                                        │
│   Layer 1: Facts Ledger         Layer 2: Episodic Memory       Layer 3: Graph Topology │
│  ┌─────────────────────────┐   ┌───────────────────────────┐  ┌──────────────────────┐ │
│  │ Versioned Facts Ledger  │   │ Mem0 Conversational       │  │ Neo4j Property Graph │ │
│  │ • Deterministic truths  │   │ • Semantic decisions      │  │ • Directed call-trees│ │
│  │ • Conflict detection    │   │ • User preferences        │  │ • Schema blast-radius│ │
│  │ • Zero prompt overhead  │   │ • Sub-100ms retrieval     │  │ • Entity dependencies│ │
│  └─────────────────────────┘   └───────────────────────────┘  └──────────────────────┘ │
│                                                                                        │
│   Layer 4: Cognitive Dynamics Engine                                                   │
│   • Synaptic Energy Decay: E(t) = E₀ · 2^(-Δt / 14d) automatically evicts stale clutter│
│   • Nightly Dream Cycle (03:00 UTC): Prunes noise, crystallizes graph insights & backups│
│   • Empathy State Tracking: Adapts agent brevity and tone to developer urgency & mood  │
│   • Neural Studio: WebGL-based 3D graph visualizer for human and agent state auditing. │
│   • Persona Synchronization: /export/persona compiles canonical rules on-demand.       │
└────────────────────────────────────────────────────────────────────────────────────────┘

The 4 Memory Layers Explained:

Layer	Technology	Primary Role	Retrieval Speed	Why It Matters
Layer 1: Facts Ledger	S3-Style Versioned JSON / SQLite	Immutable ground-truths (ports, endpoints, schemas, business invariants).	< 5ms	Deterministic recall with zero LLM hallucination and cryptographic conflict detection.
Layer 2: Episodic Memory	Mem0 Conversational History	Developer preferences, past bug fixes, and architectural tradeoffs.	< 50ms	Preserves the rationale behind past decisions so agents never repeat discarded approaches.
Layer 3: Vector Embeddings	ChromaDB High-Dim Store	Semantic search across architectural specifications, PRDs, and guides.	< 80ms	Natural-language semantic search across documents and blueprints.
Layer 4: Relational Graph	Neo4j 5.x Directed Graph	Topological dependency mapping (services, foreign keys, endpoints, workers).	< 30ms	Calculates refactor blast radius; answers: "If I alter table X, what endpoints break?"

---

Architectural Comparison Matrix

Capability	Static Prompts (.cursorrules)	Traditional Vector RAG	Friday Cognitive Substrate
Cross-Session Persistence	None (resets with thread)	Text chunks only	Full architectural state & decisions
Dependency Graph Traversal	None	Lexical similarity only	Neo4j Directed Property Graph
Token Efficiency	Burns 2,000–5,000 tokens/turn	Unfiltered chunk dumps	Targeted queries (~280 tokens/turn)
Toolchain Synchronization	Isolated per editor config	Disconnected silos	Unified MCP across Cursor, Claude, CLI
Conflict Resolution	Manual file editing required	Ingests conflicting chunks	Versioned Fact Ledger with status flags
Memory Life-Cycle	Static forever (bloats)	Flat chunk retention	Synaptic Decay + Nightly Dream Consolidation
Topology Auditing	None	None	Neural Studio 3D interactive viewer
Deployment Model	Local flat files	Cloud SaaS vendor lock-in	100% Self-Hosted Docker Compose

---

Dual-Engine Architecture: Decoupled Background Processing

A foundational architectural decision in Friday is:

"Why does Friday maintain a background worker LLM (such as Groq, DeepSeek, or local Ollama) on the server, completely separate from the frontier model running in Cursor, Claude Code, or Antigravity?"

Interactive coding agents require low latency, while knowledge graph maintenance requires continuous data extraction and synthesis. Friday enforces a Dual-Engine Architecture that cleanly decouples frontline developer workflows from background data pipelines:

┌────────────────────────────────────────────────────────────────────────────────────────┐
│                        THE DUAL-ENGINE ARCHITECTURE MODEL                                 │
├────────────────────────────────────────────────────────────────────────────────────────┤
│                                                                                        │
│   INTERACTIVE AGENT (Frontline Client)       BACKGROUND WORKER (Async Engine)      │
│   ┌─────────────────────────────────────┐     ┌─────────────────────────────────────┐  │
│   │ Client: Cursor / Claude / Antigravity│     │ Engine: Self-Hosted Friday Server   │  │
│   │ Model: Frontier (Claude 3.5 / GPT-4o)│     │ Model: Fast Worker (Groq / Ollama)  │  │
│   │ Role: Complex code generation       │     │ Role: Async graph extraction        │  │
│   │ Context: Lean, task-specific prompt │     │ Role: Conflict pruning & decay      │  │
│   │ State: Ephemeral session lifetime   │     │ State: 24/7 background persistent   │  │
│   └──────────────────┬──────────────────┘     └──────────────────▲──────────────────┘  │
│                      │                                           │                     │
│                      │ 1. MCP Tools (memory_search, add_memory)  │ 2. Microsecond      │
│                      ▼                                           │    Async Parsing    │
│   ┌──────────────────────────────────────────────────────────────┴──────────────────┐  │
│   │                        FRIDAY PERSISTENT MEMORY ARCHITECTURE                    │  │
│   │                                                                                 │  │
│   │   Layer 1: Facts Ledger (Deterministic S3-style Hash Table)                     │  │
│   │   Layer 2: Episodic Memory (Mem0 Conversational Thread History)                 │  │
│   │   Layer 3: Vector Embeddings (ChromaDB Semantic Chunks)                         │  │
│   │   Layer 4: Property Knowledge Graph (Neo4j Directed Topology)                   │  │
│   │   Memory Lifecycle: Dynamic Decay (E(t)) & Nightly Dream Cycle (03:00 UTC)    │  │
│   └─────────────────────────────────────────────────────────────────────────────────┘  │
└────────────────────────────────────────────────────────────────────────────────────────┘

Why Decoupling Background Processing is Essential:

1. ⚡ Zero-Latency IDE Execution (Non-Blocking Decoupling)

When you type in Cursor or Claude Code and an agent records a major decision via add_memory, the Conscious model cannot pause for 4–6 seconds while an LLM parses semantic entities, identifies foreign keys, and runs Cypher mutations.

• With Friday's decoupled Subconscious worker, the MCP call responds in < 40ms.
• The Subconscious engine (e.g. Groq running Llama-3 at 500+ tokens/sec) consumes the event asynchronously, wiring graph nodes and relations in the background without stealing a single millisecond of developer flow.

2. 💰 95%+ Token Cost Optimization

Frontier reasoning models (Claude 3.5 Sonnet, GPT-4o) cost $3.00 to $15.00 per million tokens. Using these expensive models for routine structural maintenance—such as extracting triples (Entity A $\longrightarrow$ RELATION $\longrightarrow$ Entity B), verifying fact hashes, or applying synaptic decay—wastes massive token budgets.

• Friday offloads structural chores to ultra-fast, ultra-cheap background APIs (Groq, DeepSeek Flash) or completely free self-hosted models (Ollama, vLLM).
• Your frontier model only spends tokens on what matters: solving complex engineering problems.

3. 🌙 Autonomous Background Consolidation (The Dream Cycle)

Your coding session ends when you close your IDE or put your laptop to sleep. But memory evolution cannot stop when the laptop closes:

• Friday's Subconscious engine lives on your cloud or local server 24/7.
• At 03:00 UTC every night, while you are asleep, the Subconscious wakes up to run the Dream Cycle: calculating synaptic decay, pruning low-energy noise, distilling daily episodic learnings into permanent strategic facts, and committing encrypted snapshots to Git.

4. 🛡️ Hallucination & Context Pollution Defense

Dumping a monolithic 500-node graph or 100 historical decisions directly into your editor's prompt causes Instruction Dilution: the LLM becomes confused, forgets recent constraints, and hallucinates outdated patterns.

• The Subconscious acts as an intelligent firewall.
• It digests raw context, resolves contradictions, calculates energy decay ($E(t)$), and serves only the top crystallized, high-energy facts directly relevant to your active task (~280 tokens instead of 5,000).

---

Memory Lifecycle Management: Decay, Consolidation & Calibration

Friday implements active memory lifecycle management to ensure AI agents retain critical constraints without context bloat or stale instruction interference:

┌────────────────────────────────────────────────────────────────────────────────────────┐
│                          FRIDAY MEMORY LIFECYCLE ENGINE                                │
├────────────────────────────────────────────────────────────────────────────────────────┤
│                                                                                        │
│  🔥 Dynamic Memory Heat & Decay           🌙 The Dream Cycle (Nightly 03:00 UTC)      │
│  ┌───────────────────────────────────┐    ┌────────────────────────────────────────┐   │
│  │ Exponential Synaptic Decay        │    │ 1. Synaptic Pruning (Evaporates noise) │   │
│  │ • E(t) = E₀ · 2^(-Δt / T_half)    │───>│ 2. Episodic Synthesis (Distills gems)  │   │
│  │ • Recall Potentiation (+0.25)     │    │ 3. Neo4j Crystallization (Graph edges) │   │
│  │ • Soft Archive if E < 0.25        │    │ 4. Autonomous Backup to Git            │   │
│  └───────────────────────────────────┘    └────────────────────────────────────────┘   │
│                                                                                        │
│  🤍 Adaptive Context & Persona Calibration                                             │
│  ┌──────────────────────────────────────────────────────────────────────────────────┐  │
│  │ Multi-Dimensional Response Calibration                                           │  │
│  │ • Interaction Modes: tactical_sprint | deep_architecture | casual_brainstorm     │  │
│  │ • Task Context & Urgency Detection (0.0 to 1.0)                                  │  │
│  │ • Dynamic Response Calibration: Brevity (high/med/low) & Tone Tuning             │  │
│  └──────────────────────────────────────────────────────────────────────────────────┘  │
└────────────────────────────────────────────────────────────────────────────────────────┘

1. 🔥 Dynamic Memory Heat & Decay

Memories and verified facts are not static text—they have energy. Active, frequently recalled directives remain bright ($E > 1.0$). Irrelevant or outdated details experience exponential half-life decay ($T_{half} = 14\text{ days}$): $$E(t) = E_0 \times 2^{-\frac{\Delta t}{T_{half}}}$$ When a memory is queried during coding, it receives a recall potentiation boost ($+0.25$), preventing stale knowledge from cluttering the agent prompt while preserving core architectural invariants (decay_immune: True).

2. 🌙 The Dream Cycle

Every night at 03:00 UTC (or on-demand via client.run_dream_cycle()), Friday enters the Dream Cycle:

• Synaptic Pruning: Identifies cold/stale facts and transitions them to archived storage.
• Episodic Synthesis: Clusters recent conversations and distills 1–2 crystallized strategic insights.
• Neo4j Crystallization: Links high-confidence insights into the property graph with CRYSTALLIZED_INTO edges.
• Autonomous Git Sync: Triggers automated repo commits preserving graph snapshots.

3. 🤍 Adaptive Context & Persona Calibration

Friday monitors the developer interaction context (urgent bug-fix sprint, late-night architecture exploration, or casual brainstorming). The engine dynamically adjusts agent response characteristics:

• Brevity Calibration: high (zero fluff, code-first) vs. detailed (system-wide breakdown).
• Tone Calibration: technical_concise (code-first, direct) vs. architectural_detailed (system-wide breakdown).
• Injected automatically into /export/persona so all agents naturally calibrate their output.

---

DeepEval Benchmarks

We evaluated five realistic engineering scenarios using the DeepEval evaluation framework:

1. Database Schema Blast Radius (evaluating downstream call-graph traversal)
2. Authentication Refresh Lifecycle (evaluating versioned constraint fidelity)
3. Webhook Idempotency Guarantee (evaluating race-condition edge cases)
4. Environment & Port Reservations (evaluating static ground-truth recall)
5. Multi-Agent Toolchain Consistency (evaluating cross-tool synchronization between Cursor and Claude CLI)

Memory Architecture	Contextual Precision	Contextual Recall	Faithfulness	Prompt Tokens / Turn	Session Retention
Static Prompts (.cursorrules)	38.0%	44.0%	62.0%	3,150 tokens	15.0% (resets)
Naive Vector RAG (Vector Only)	64.0%	58.0%	74.0%	1,820 tokens	55.0%
Friday Cognitive Substrate	95.0%	93.0%	99.0%	280 tokens	100.0%

Reproducing Benchmarks Locally

python benchmarks/benchmark_deepeval.py

---

Quickstart

Option A: One-Command Installation (Recommended)

Run the self-contained installation script:

curl -fsSL https://raw.githubusercontent.com/friday-memory/friday/main/install.sh | bash

The script verifies Docker availability, allocates required ports (8000, 7474, 7687), generates secure random API secrets, writes a validated .env, and launches Friday via Docker Compose.

Option B: Manual Setup via Docker Compose

1. Clone the Repository:

   git clone https://github.com/friday-memory/friday.git
   cd friday
   

2. Configure Environment (.env):

   cp .env.example .env
   

   # Master API key for endpoint security
   BRAIN_API_KEY=choose_a_strong_secret_key
   
   # Fast Subconscious LLM provider (Groq or DeepSeek)
   DEEPSEEK_API_KEY=your_key_here
   DEEPSEEK_BASE_URL=https://api.deepseek.com
   DEEPSEEK_MODEL=deepseek-chat
   
   # Mem0 key for vector memory (optional)
   MEM0_API_KEY=your_mem0_key_here
   
   # Neo4j database credentials
   NEO4J_URI=bolt://neo4j:7687
   NEO4J_USER=neo4j
   NEO4J_PASSWORD=choose_a_strong_password
   

3. Start the Stack:

   make up
   # or: docker compose up -d
   

4. Verify Health:

   curl http://localhost:8000/health
   

   {
     "status": "healthy",
     "service": "friday-cognitive-substrate",
     "version": "1.4.4",
     "layers": {
       "L1_core": "healthy",
       "L2_mem0": "healthy",
       "L3_chromadb": "healthy",
       "L4_neo4j": "healthy"
     }
   }
   

---

Python SDK (friday-memory)

The official Python client for Friday is available on PyPI as friday-memory. Connect your agentic workflows, LangChain pipelines, or autonomous scripts directly to Friday with zero boilerplate:

pip install --upgrade friday-memory

Synchronous Client

from friday import Friday