NeuralMemory MCP
Agent memory as a graph of linked facts: decisions and errors outlive the session, with no vector database or paid API calls
- 243 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Plugin hooks run commands at session start and stop and after every tool call, and memory is written to disk
- Cloud sync and Telegram backup, if enabled, send the memory contents to external services
Install
- Works in
Manual install
pip install neural-memoryInstalls the `nmem` CLI and the `nmem-mcp` MCP server. The server initializes itself after the client restarts.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Plugin hooks run commands at session start and stop and after every tool call, and memory is written to disk
- Cloud sync and Telegram backup, if enabled, send the memory contents to external services
This is third-party code. Review the repository files before installing.
What it does
NeuralMemory stores an agent's experience as linked neurons and retrieves it through spreading activation: a query lights up related concepts until the right memory surfaces. Relations are named (`CAUSED_BY`, `LEADS_TO`, `RESOLVED_BY`, `CONTRADICTS`), so a question like "why did the service fail on Tuesday" follows a chain of causes. No embeddings are needed and the free tier runs fully locally. The `nmem-mcp` MCP server exposes 63 tools, but three are enough to start: `nmem_remember` saves a record and detects its type, tags and links, `nmem_recall` retrieves records, and `nmem_health` grades the memory from A to F. The `nmem` CLI (82 commands) sits beside it: `nmem remember`, `nmem recall`, `nmem brain export` and `nmem serve` for a web dashboard with a graph and a timeline. Memory has fourteen types (fact, decision, error, preference, instruction and more), decays and strengthens over time, and episodes consolidate into lasting knowledge. You can train it on PDF, DOCX, PPTX, HTML, XLSX and CSV files or migrate memory from ChromaDB, Mem0, Cognee and LlamaIndex. The Claude Code plugin adds three skills and session hooks.
Who it is for. Developers whose agent in Claude Code or Cursor starts every session from zero and forgets the decisions, bugs and agreements of the project.
Good fit when
- An architecture decision was made a week ago and the agent advises the opposite in a new session
- A failure keeps coming back and you need the chain of causes from symptom to the original change
- Memory piled up in Mem0 or ChromaDB and you want to move it without losing records
- Memory has to live on several computers but only inside your own cloud account: sync runs through your own Cloudflare Worker
- Project documents in PDF, DOCX and spreadsheets should become permanent agent knowledge
Not a fit when
- A `CLAUDE.md` or `AGENTS.md` file with a dozen rules is enough, and a separate memory server is overkill
- You need semantic search over hundreds of thousands of records: it ships in the paid Pro tier, while the free tier matches keywords
- The machine runs Python below 3.11, which the package does not accept
Example request
Recall why we dropped Redis in the queue service, and save the final database choice as a decision tagged architectureMCP
- Transport
- stdio
- Authentication
- not required
Limitations
- Python 3.11 or newer is required.
- Semantic search, tiered compression and automatic sync belong to the paid Pro tier at 9 dollars a month. The free tier matches keywords and has no cap on the number of records.
- Syncing between machines needs your own Cloudflare account and a Worker deployed there.
- The Claude Code plugin hooks run `nmem-hook-*` commands at session start, before context compaction, on stop and after each tool call. Hook timeouts range from 10 to 30 seconds.
How to disable. In Claude Code disable the plugin through `/plugin`, and remove the `neural-memory` entry from the MCP configuration in other clients. Uninstall the package with `pip uninstall neural-memory`. Local memory databases are stored separately and are deleted by hand.
FAQ
How do I install NeuralMemory in Claude Code?
Run `/plugin marketplace add nhadaututtheky/neural-memory`, then `/plugin install neural-memory@neural-memory-marketplace`. The plugin wires in the MCP server, three skills and session hooks.
How do I connect NeuralMemory to Cursor or Windsurf?
Install the package with `pip install neural-memory` and add a `neural-memory` server with the command `nmem-mcp` to the editor's MCP configuration. It initializes on first use, so no `init` command is needed.
Does it need embeddings or API keys?
No. The free tier works without LLM calls and without embeddings, and retrieval goes through the link graph and keywords. Local and cloud embedding models are separate package extras.
How is the paid Pro tier different from the free one?
Pro switches storage to InfinityDB and adds semantic search, five-tier compression and automatic sync. The license is activated with `nmem shared activate`, and every free tool keeps working.
Where is the data stored and how do I move it to another computer?
Memory lives in a local SQLite database by default. To move it, `nmem sync` synchronizes through a Worker in your own Cloudflare account, and `nmem brain export -o backup.json` writes a copy to a file.
Discussion
No questions or reviews yet.
Related
MCP serversSelf-hosted long-term memory for agents: semantic search, a knowledge graph and autonomous consolidation over MCP and a REST API
Long-term memory for Claude Code, Cursor and Codex: the agent saves decisions and finds them again in later sessions
A self-hosted MCP memory server that keeps preferences and decisions across conversations in Claude Code, Cursor and other clients
Long-term agent memory in local Markdown files with a link graph, semantic search and MCP access
A long-term memory MCP server for AI agents: facts fade along a forgetting curve and related records are compressed into a summary
A long-term memory plugin that keeps decisions, preferences and task state across sessions and injects only what a query needs
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| NeuralMemory (this card) | no | not required | 243 | |
| Shodh-Memory | no | key is optional | 322 | a single Rust binary that runs offline and never calls a language model; NeuralMemory builds a causal link graph with record decay |
| Basic Memory | no | not required | 4,495 | keeps memory in local Markdown files with links between notes; NeuralMemory keeps a neuron graph in SQLite |
| MCP Memory Service | no | not required | 2,090 | self-hosted memory with semantic search and a REST API; NeuralMemory searches a link graph without embeddings |
README in short
The documentation opens with the idea of a graph in place of vector search and a comparison table against RAG. It continues with cloud sync, a feature list, command and Python examples, the Pro tier, setup per client and a speed comparison with Mem0 and Cognee.