About
GnoSys solves one of the biggest limitations of AI agents — the inability to remember anything between sessions. It acts as a centralized "brain" for LLMs, backed by a single SQLite database (~/.gnosys/gnosys.db) shared across all projects with sub-10ms reads and automatic backups. The platform provides 50+ MCP tools covering memory creation, retrieval, audit logging, and rollback. Its three-scope architecture (project, user, global) lets memories be scoped appropriately — project context stays with the codebase, user preferences follow the developer everywhere, and global knowledge spans an entire organization. Federated search applies intelligent tier boosting (current project 1.8×, user 1.0×, global 0.7×) with recency and reinforcement signals, so the most relevant memories surface first. Memory Lensing allows compound filtering by category, tags, confidence, author, and dates. Dream Mode runs idle-time consolidation — decaying stale confidence scores, writing self-critiques, generating summaries, and discovering relationships between memories — analogous to biological sleep. Multimodal ingestion supports PDFs, images (OCR), audio, video, and DOCX files. A Web Knowledge Base feature converts any website into a searchable markdown directory. GnoSys is ideal for AI-assisted developers who want coding agents to retain context across long-running projects, teams building knowledge-management pipelines for LLM agents, and anyone who needs a self-hosted, auditable memory layer without relying on proprietary vector databases.
Key Features
- Centralized SQLite Brain: One ~/.gnosys/gnosys.db file shared across all projects with a 6-table schema, sub-10ms reads, automatic backups, and a full audit log with non-destructive rollback.
- Federated Search with Tier Boosting: Cross-scope search ranks results by proximity (current project 1.8×, user 1.0×, global 0.7×), recency, and reinforcement signals to surface the most relevant memories first.
- Dream Mode Consolidation: Idle-time background process that decays stale confidence scores, writes self-critiques, generates summaries, and discovers relationships between memories — like biological sleep for your knowledge base.
- Multimodal Ingestion: Ingest PDFs, images (OCR), audio, video, and DOCX files as structured memories. LLM-powered extraction automatically generates titles, tags, categories, and keyword clouds.
- Obsidian Export & Wikilinks: Export any memory or the entire database to Obsidian-compatible markdown with [[wikilinks]] and backlinks via a single command, building a navigable knowledge graph.
Use Cases
- Giving AI coding assistants like Claude Code or Cursor persistent project memory so they remember architecture decisions, conventions, and past debugging sessions across restarts.
- Building a personal or team knowledge base that AI agents can query, update, and consolidate over time without relying on external vector databases.
- Tracking multi-project status and production-readiness scores via the portfolio dashboard, with AI agents writing structured status snapshots through guided update workflows.
- Ingesting company documents (PDFs, DOCX, audio transcripts) into a structured, searchable memory store that LLMs can reference during conversations.
- Exporting agent-generated knowledge to Obsidian for human review, creating a linked knowledge graph with wikilinks and backlinks that bridges AI and personal note-taking.
Pros
- Fully Open Source (MIT): Self-hosted with no vendor lock-in, no subscription fees, and full visibility into how memories are stored, indexed, and retrieved.
- No Vector Database Required: Hybrid search and tier boosting are implemented directly on SQLite, eliminating the complexity and cost of maintaining a separate vector store.
- Broad MCP Client Compatibility: Works out of the box with Claude Desktop, Claude Code, Cursor, Codex, and Gemini CLI via a single gnosys init <ide> command.
- Rich Memory Tooling: 50+ MCP tools covering scoped memory, audit logs, process tracing, portfolio dashboards, bulk import, and multi-machine sync go far beyond basic key-value storage.
Cons
- Developer-Focused Setup: Installation requires Node.js and npm, and advanced features like the sandbox runtime or bulk import involve CLI commands that may be unfamiliar to non-technical users.
- Local-First Architecture: The default SQLite-based design is optimized for single-machine or NAS-shared use; teams needing real-time cloud sync across many users will require additional infrastructure.
- Early-Stage Ecosystem: As a newer open-source project, documentation and community support are still maturing compared to established memory or RAG frameworks.
Frequently Asked Questions
What is GnoSys and who is it for?
GnoSys is an open-source MCP (Model Context Protocol) server that gives AI agents persistent memory stored in a local SQLite database. It is primarily aimed at developers who use AI coding assistants like Claude Code or Cursor and want those agents to retain context, preferences, and project knowledge across sessions.
How do I connect GnoSys to my AI client?
Install GnoSys globally with `npm install -g gnosys`, then run `gnosys init <ide>` from your project directory (e.g., `gnosys init claude` or `gnosys init cursor`). GnoSys automatically writes the MCP configuration file for the selected client.
What is Dream Mode?
Dream Mode is an idle-time background consolidation process that mimics biological sleep. When your agent is inactive, it decays confidence scores on stale memories, runs self-critiques, generates summaries, and discovers relationships between stored memories to keep the knowledge base healthy and relevant.
Does GnoSys require a vector database?
No. GnoSys performs hybrid search — combining keyword and semantic-style retrieval — directly on its SQLite database using tier boosting, recency signals, and reinforcement weighting. No separate vector database or embedding service is required.
Can I use GnoSys across multiple machines?
Yes. GnoSys supports multi-machine sync by sharing the gnosys.db file over a NAS or shared drive. It maintains a local cache for speed, treats the remote copy as the source of truth, and includes built-in conflict detection with AI-mediated resolution.
