# Friday: AI memory that actually sticks

> Open-source cognitive memory layer stops AI coding agents from forgetting your architecture between sessions.
- Title: Friday: AI memory that actually sticks · GitHub | Shortify
- Summary: Open-source cognitive memory layer stops AI coding agents from forgetting your architecture between sessions. AI agents need persistent external memory…
- Keywords: friday, ai coding agents, persistent memory, cursor, claude, copilot, model context protocol, neo4j, knowledge graph, self-hosted, open source, cognitive architecture
- Source: GitHub — https://github.com/itskie/friday
- Read time: 3 min
- Topics: friday, ai coding agents, persistent memory, cursor, claude, copilot, model context protocol, neo4j, knowledge graph, self-hosted, open source, cognitive architecture
## AI agents forget everything
Cursor, Claude, and Copilot lose all context when you close a chat or restart. Developers hit the same architectural problems repeatedly because the AI has no persistent memory between sessions.
## Prompt files don't scale
Massive .cursorrules files burn 2,000-3,000 tokens on every interaction, even for simple fixes. Rules rot over time, contradict each other, and don't sync across your team's tools.

> The Token Tax: You burn 2,000, 3,000 prompt tokens on every interaction, even when merely asking the model to fix CSS padding.
## Vector search misses dependencies
Standard RAG retrieves files containing matching text but cannot traverse dependency graphs. It cannot predict what breaks when you rename a database column across multiple services.

> Software architectures are directed graphs, not flat text documents. Without a relational graph database, agents cannot predict the blast radius of structural changes.
## Friday solves session amnesia
An open-source, self-hosted cognitive memory layer that persists architecture decisions, schemas, and preferences across sessions via the Model Context Protocol.
## Three-layer memory architecture
Facts Ledger stores versioned ground truth with zero token cost. Episodic Memory tracks decisions chronologically. Graph Topology maps dependencies as a Neo4j knowledge graph.
## Auto-graph engine wires knowledge
Every memory you store is automatically analyzed by an LLM to extract entities and relationships, which are wired into Neo4j without manual tagging or YAML configuration.
## Neural Studio visualizes your brain
A 3D WebGL interface maps your knowledge graph onto an anatomical cerebral cortex with live synaptic action potentials, entity inspection, and full-text search.
## Benchmarks show 95% precision
Friday achieves 95% contextual precision, 93% recall, and 99% faithfulness while using only 280 tokens per turn versus 3,150 for static prompts.

> 91% Token Reduction: Friday injects targeted snippets (~280 tokens) instead of burning 3,000+ tokens of raw prompt files on every keystroke.
## Blast radius capture is 96% accurate
The Neo4j property graph captures foreign keys, service handlers, and API consumers that text-matching vector embeddings miss entirely.

> 96% Blast-Radius Capture: The Neo4j property graph captures foreign keys, service handlers, and API consumers that text-matching vector embeddings miss.
## One-command setup with Docker
Run a single bash script that verifies Docker, provisions containers, generates secure keys, and prints ready-to-use MCP configurations for Cursor and Claude.
## Connect all your AI tools
Friday syncs across Cursor, Claude Code, Antigravity IDE, VS Code, and custom agents via the Model Context Protocol. One brain, all your tools.

> A shared brain superpower: An architectural rule stored by Claude Code in your terminal is immediately accessible to Cursor, Antigravity IDE, or Codex on your desktop. Zero manual syncing.
## Four native MCP tools
get_context loads all facts at session start. memory_search retrieves relevant context. add_memory persists decisions. add_fact stores atomic rules.
## Versioned facts prevent conflicts
Discrete facts are stored with immutable version history. Old versions are superseded, never deleted, creating a full audit trail of what changed and when.
## Dynamic persona export syncs rules
Pull canonical rules directly from Friday's central brain to generate .cursorrules or AGENTS.md files that stay in sync across your entire team.
## REST API for any agent
Store and retrieve memories via HTTP without MCP. Any custom agent, CI loop, or script can interact with Friday using simple curl commands.
## Semantic search via ChromaDB
Instead of dumping entire memory into every prompt, Friday uses vector search to retrieve only the most relevant context for each query.
## Self-hosted, not vendor lock-in
Deploy on your own infrastructure with Docker Compose. No cloud SaaS dependency. Full control over your team's architectural knowledge.
## Roadmap includes SDKs and CLI
Version 1.1 adds Python and TypeScript SDKs plus a friday CLI for terminal commands. Version 1.2 brings GitHub Actions, Slack, and Jira integrations.
## Built for AI-native developers
Friday is open source, MIT licensed, and built by developers tired of repeating themselves to their AI. Contributions welcome.

> Made with ❤️ by developers who were tired of repeating themselves to their AI.
## Key takeaway

AI agents need persistent external memory organized as knowledge graphs, not just prompt files or vector search, to understand software dependencies and avoid repeated mistakes.