Memory Backends

The FL Repos MCP server can be paired with a memory backend to give AI agents persistent, searchable memory across sessions. This allows the agent to remember previous analyses, team decisions, and contextual notes about the codebase. Two memory backends are supported: Graphiti (the default and recommended option) and Zep (an alternative).

Graphiti

Graphiti is a temporal knowledge graph built on top of Neo4j/FalkorDB. It stores memory as episodes (timestamped facts and observations) and automatically extracts entities and relationships from natural language, weaving them into a knowledge graph alongside the FL Repos data.

Deployment

The Graphiti server is deployed at:

https://graphiti-production-3043.up.railway.app

Pre-loaded episodes

The Graphiti instance for FL Repos has 641 episodes already ingested. These include:

Repository analysis summaries from previous AI sessions
Architectural observations and cross-repo patterns
Team-added notes about specific services and components
Historical context about why certain design decisions were made

Configuration

Set the GRAPHITI_URL environment variable in your MCP server config:

{
  "mcpServers": {
    "fl-repos": {
      "command": "python",
      "args": ["-m", "intelligence.mcp.server"],
      "env": {
        "FL_API_URL": "https://api.agentic.forwardlane.com",
        "FL_API_TOKEN": "YOUR_API_TOKEN_HERE",
        "GRAPHITI_URL": "https://graphiti-production-3043.up.railway.app"
      }
    }
  }
}

export FL_API_URL="https://api.agentic.forwardlane.com"
export FL_API_TOKEN="YOUR_API_TOKEN_HERE"
export GRAPHITI_URL="https://graphiti-production-3043.up.railway.app"

How Graphiti stores memory

When the AI agent calls add_memory, Graphiti:

Parses the episode content for named entities (repos, functions, contributors, etc.)
Links extracted entities to existing nodes in the knowledge graph
Timestamps the episode for temporal reasoning (“what did we know last week?”)
Indexes the content for semantic similarity search

When the agent calls search_memory, Graphiti:

Embeds the search query
Searches both the graph structure (entity links) and vector index (semantic similarity)
Ranks results by relevance and recency
Returns the most relevant episodes with their timestamps and extracted entities

Temporal reasoning

Graphiti’s temporal model means the agent can reason about how knowledge changes over time:

"When was the decision to split fl-auth-service from the backend first recorded?"
"What has changed about the billing service since last month?"
"Show me the most recent observations about the data pipeline."

Zep

Zep is an episodic memory service designed for AI applications. It stores sessions, messages, and facts, and provides semantic search over them.

When to use Zep

Use Zep instead of Graphiti when:

You prefer a managed cloud service with a dashboard
You want tight integration with LangChain or LlamaIndex
You need session-level memory (conversation history) rather than graph-linked facts

Configuration

export ZEP_API_URL="https://your-zep-instance.api.getzep.com"
export ZEP_API_KEY="your-zep-api-key"

{
  "mcpServers": {
    "fl-repos": {
      "command": "python",
      "args": ["-m", "intelligence.mcp.server"],
      "env": {
        "FL_API_URL": "https://api.agentic.forwardlane.com",
        "FL_API_TOKEN": "YOUR_API_TOKEN_HERE",
        "ZEP_API_URL": "https://your-zep-instance.api.getzep.com",
        "ZEP_API_KEY": "your-zep-api-key"
      }
    }
  }
}

If both GRAPHITI_URL and ZEP_API_URL are set, Graphiti takes precedence. To use Zep, set only the Zep variables.

Backend Comparison

Feature	Graphiti	Zep
Storage model	Temporal knowledge graph	Episodic sessions
Entity extraction	Automatic	Manual
Graph linking	Yes — links to FL Repos nodes	No
Temporal queries	Yes	Limited
Semantic search	Yes	Yes
Pre-loaded data	641 episodes	—
Managed hosting	Yes (Railway)	Yes (getzep.com)
LangChain integration	No	Yes

Using Memory in Practice

Adding a note about a repository

You: Note that the fl-portfolio-service needs to be refactored to use the new auth library before Q2.

Claude: [calls add_memory with the content]
Memory stored: "fl-portfolio-service refactor needed — migrate to new auth library before Q2"

Searching for past decisions

You: What have we noted about the authentication service previously?

Claude: [calls search_memory("authentication service")]
Found 3 relevant episodes:
1. (2024-10-15) "fl-auth-service was split from forwardlane-backend in Sprint 42..."
2. (2024-11-02) "Auth service now owns JWT validation; backend delegates all token checks..."
3. (2024-11-20) "Decision: fl-auth-service will adopt OAuth2 device flow for CLI tools..."

Combining memory with graph queries

The most powerful pattern is combining live graph data with stored memory context:

You: Based on what we know about the billing service and current code analysis,
     what are the highest-risk areas?

Claude:
1. [calls search_memory("billing service")] — retrieves past notes
2. [calls get_repo_detail("fl-billing")] — gets current stats
3. [calls list_security(repo="fl-billing")] — gets security findings
Synthesizes: "Based on 2 past notes and current analysis: 3 high-severity security
findings in fl-billing/payment_processor.py, plus a previously noted tech debt item
around the webhook retry logic..."

Get Started

Products

API Reference

CLI

FL Web

Python SDK

MCP Server

Graph Schema

Memory Backends

Memory Backends

Graphiti

Deployment

Pre-loaded episodes

Configuration

How Graphiti stores memory

Temporal reasoning

Zep

When to use Zep

Configuration

Backend Comparison

Using Memory in Practice

Adding a note about a repository

Searching for past decisions

Combining memory with graph queries

​Memory Backends

​Graphiti

​Deployment

​Pre-loaded episodes

​Configuration

​How Graphiti stores memory

​Temporal reasoning

​Zep

​When to use Zep

​Configuration

​Backend Comparison

​Using Memory in Practice

​Adding a note about a repository

​Searching for past decisions

​Combining memory with graph queries

Memory Backends

Graphiti

Deployment

Pre-loaded episodes

Configuration

How Graphiti stores memory

Temporal reasoning

Zep

When to use Zep

Configuration

Backend Comparison

Using Memory in Practice

Adding a note about a repository

Searching for past decisions

Combining memory with graph queries