Simpler setup
One entry in your client's MCP configuration starts it with npx or Docker. Okto Neuron runs an installer, a daemon and a UI.
Okto Neuron vs the MCP Knowledge Graph Memory server. Who writes the graph.
The Knowledge Graph Memory server is one of the Model Context Protocol reference servers. Your agent writes entities, relations and observations into it through MCP tools, and it keeps them in a local JSONL file. Okto Neuron builds its graph from Markdown notes you write, with an extraction LLM and a confidence gate. This page marks where each fits.
- 5
- MCP tools: ask, explore, remember, list_vaults, init_vault
- 0
- Hosted tiers. Runs on your machine only
- LoCoMo
- Measured with our own harness. Results on the benchmarks page
- 0.3.0
- Prerelease. Claude Code is the wired MCP client
The short answer.
The Knowledge Graph Memory server from the Model Context Protocol reference servers is a small local store: your agent creates entities, relations and observations through nine tools, and the server keeps them in a JSONL file. Okto Neuron builds its graph from your Markdown notes with an extraction LLM and a confidence gate, and keeps a byte-range pointer on every claim.
Pick the reference server if you want the simplest possible memory, started by your client with npx or Docker, with no LLM of its own.
Pick Okto Neuron if you want the graph derived from notes you already write, with citations back to the source.
01 · The table
Side by side, row by row.
Reference-server cells paraphrase its README, read on 2026-09-26. Okto Neuron cells were checked against the 0.3.0 source on 2026-09-26.
| Dimension | MCP memory server | Okto Neuron |
|---|---|---|
| Who writes facts | The agent, through tool calls that create entities, relations and observations | An extraction LLM reads your notes on ingest; entities commit at confidence 0.75 or higher, claims pass a separate grounding gate, and anything held back waits for review |
| Storage | A JSONL file, set with MEMORY_FILE_PATH (default memory.jsonl in the server directory) | Okto Grafx by default (embedded), derived from the Markdown vault and rebuildable from it |
| Provenance | Observations are strings attached to an entity. No pointer back to a source file is described | Each claim points to a block: vault-relative path, byte_start, byte_end and a sha256 content_hash |
| Tools | 9: create_entities, create_relations, add_observations, delete_entities, delete_observations, delete_relations, read_graph, search_nodes, open_nodes | 5: ask, explore, remember, list_vaults, init_vault |
| How it runs | The client launches it as a local npx or docker command, per the README's configuration examples | One daemon started by the installer, reached over HTTP on 127.0.0.1:8201 with a bearer token |
| LLM needed | No | Yes, for ingest and for written answers. explore makes no LLM call |
| License | MIT | Elastic License 2.0 from 0.3.0 (source-available). Releases up to 0.2.0 stay Apache 2.0 |
| Maturity | A reference implementation published on npm as @modelcontextprotocol/server-memory | Prerelease 0.3.0. The upgrade from 0.2.0 has been rehearsed on Linux with a locally built wheel; macOS and Windows are not yet verified on 0.3.0 |
02 · Where the reference server is ahead
What it has that Okto Neuron does not.
No model to run
It stores what the agent sends and needs no LLM of its own. Okto Neuron needs an LLM endpoint for ingest and for written answers.
An official reference implementation
It lives in the Model Context Protocol servers repository and is published on npm, under MIT.
One-click VS Code install
Its README carries install badges for VS Code and VS Code Insiders. Okto Neuron's installer wires Claude Code; other clients are not tested.
03 · Where Okto Neuron differs
Where Okto Neuron works differently.
A pointer back to the source
Each claim carries the vault-relative path, byte_start, byte_end and a sha256 content_hash of the block it was read from, returned with include_sources=true.
A failed answer looks failed
Every ask stamps retrieval.synthesis_status: ok, empty, provider_error, truncated, abnormal_stop or no_llm.
Supersession on ingest
When you replace a fact in a note, the next ingest asks an LLM correction judge whether the new claim corrects the old one. If it confirms, the old claim is dated and linked to the new one with a supersedes edge, within the same document.
The notes stay the source of truth
The graph is derived from your Markdown vault and can be deleted and rebuilt from it. A rebuild re-runs LLM extraction, so the new graph can differ from the old one, and the history of superseded claims is not recreated.
Questions about this comparison.
Do these benchmark numbers compare?
The reference server's README publishes no benchmark results. Our LoCoMo results come from our own harness and a local judge, on categories 1-4 only. They are on the benchmarks page.
Is Okto Neuron open source?
It is source-available. From 0.3.0 the license is the Elastic License 2.0 with an addendum covering SaaS, competing services, internal use and branding. Releases up to 0.2.0 stay under Apache 2.0. The source is on GitHub.
Does Okto Neuron have a hosted version?
No. It runs on your machine, binds 127.0.0.1 and needs no account.
04 · Sources
Where each claim about the reference server comes from.
Every statement about the reference server on this page comes from its README, read on 2026-09-26. If it has changed since, the page is wrong and we would like to know.
| Claim area | Primary source |
|---|---|
| Entities, relations, observations, nine tools, JSONL storage, npx and Docker setup, VS Code badges, MIT license | https://raw.githubusercontent.com/modelcontextprotocol/servers/main/src/memory/README.md |
Point it at your notes. Check the bytes yourself.
Prerelease 0.3.0. Free to run locally. No account required.