Long-term memory MCP server for agents, with URI-addressed memories, namespaces, search, aliases, glossary triggers, a visual dashboard, and reviewable rollback snapshots backed by SQLite or PostgreSQL.
Nocturne Memory can create, update, delete, alias, trigger, and search persistent memory nodes., Stdio startup can launch a local admin dashboard, initialize the database, and build the frontend if dependencies are present., Network-facing SSE or HTTP mode refuses to start without an API token when bound to a reachable host, but localhost-only deployments should still set one., Write tools can change long-term agent behavior; review snapshots and rollback queues before accepting broad edits or cleanup operations., Keep `public_readonly_mcp` enabled only for demos or intentionally read-only deployments.
Privacy notes
Memories can contain personal history, health details, credentials accidentally pasted into chats, private project context, relationship content, strategies, drafts, and agent identity instructions., Database files, PostgreSQL URLs, API tokens, config files, snapshots, dashboard views, MCP transcripts, and search results can expose the full memory graph., The README includes public demo guidance, but private memories should be stored only in a controlled self-hosted instance., Redact memory URIs, node content, boot URIs, glossary triggers, search terms, database paths, API tokens, and snapshot diffs before sharing logs or screenshots.
Author
Dataojitori
Submitted by
oktofeesh1
Claim status
unclaimed
Last verified
2026-06-06
Decision playbook
Review trust signals before you adopt
Signals are present but mixed. Use the checklist below to confirm the source and operational safety for your environment.
Compare context
Selected
0
Current score
63
Baseline
—
Delta
No baseline selected
No major trust-signal divergence detected in the current selection.
Source and provenance checks
Needs review
Confirm ownership and provenance before trusting install instructions.
Source link availableRequired
Open the canonical repository and verify ownership.
Done
Source provenance statusRequired
Marked as source-backed.
Done
Metadata reviewed
No reviewed flag detected in metadata.
Pending
Safety and privacy checks
Complete
Validate risk disclosures before installation or API wiring.
Safety notes presentRequired
Review the listed safety guidance before running commands.
Done
Privacy notes presentRequired
Review data handling notes before connecting accounts or secrets.
Done
Trust level risk gateRequired
Trust level does not block evaluation.
Done
Package and install checks
Needs review
Check package metadata and artifact integrity signals.
Install payload available
Install or copy payload is available for review.
Done
Package verification flag
No package verification flag provided.
Pending
Checksum metadata
No checksum provided for downloaded artifact.
Pending
Compare-driven decision checks
Needs review
Use compare context to validate trade-offs before adoption.
Compare tray has multiple entries
Add at least one more entry to compare trust differences.
5 safety and 4 privacy notes across 5 risk areas. Review closely: credentials & tokens, permissions & scopes.
5 areas
SafetyData retentionNocturne Memory can create, update, delete, alias, trigger, and search persistent memory nodes.
SafetyPermissions & scopesStdio startup can launch a local admin dashboard, initialize the database, and build the frontend if dependencies are present.
SafetyCredentials & tokensNetwork-facing SSE or HTTP mode refuses to start without an API token when bound to a reachable host, but localhost-only deployments should still set one.
SafetyLocal filesWrite tools can change long-term agent behavior; review snapshots and rollback queues before accepting broad edits or cleanup operations.
SafetyGeneralKeep `public_readonly_mcp` enabled only for demos or intentionally read-only deployments.
PrivacyCredentials & tokensMemories can contain personal history, health details, credentials accidentally pasted into chats, private project context, relationship content, strategies, drafts, and agent identity instructions.
PrivacyCredentials & tokensDatabase files, PostgreSQL URLs, API tokens, config files, snapshots, dashboard views, MCP transcripts, and search results can expose the full memory graph.
PrivacyData retentionThe README includes public demo guidance, but private memories should be stored only in a controlled self-hosted instance.
PrivacyCredentials & tokensRedact memory URIs, node content, boot URIs, glossary triggers, search terms, database paths, API tokens, and snapshot diffs before sharing logs or screenshots.
Disclosure: MIT-licensed open-source memory server. The project uses strong AI-persona framing in its README; this entry describes the source-backed MCP memory behavior without endorsing storing sensitive personal data unreviewed.
Safety notes
Nocturne Memory can create, update, delete, alias, trigger, and search persistent memory nodes.
Stdio startup can launch a local admin dashboard, initialize the database, and build the frontend if dependencies are present.
Network-facing SSE or HTTP mode refuses to start without an API token when bound to a reachable host, but localhost-only deployments should still set one.
Write tools can change long-term agent behavior; review snapshots and rollback queues before accepting broad edits or cleanup operations.
Keep `public_readonly_mcp` enabled only for demos or intentionally read-only deployments.
Privacy notes
Memories can contain personal history, health details, credentials accidentally pasted into chats, private project context, relationship content, strategies, drafts, and agent identity instructions.
Database files, PostgreSQL URLs, API tokens, config files, snapshots, dashboard views, MCP transcripts, and search results can expose the full memory graph.
The README includes public demo guidance, but private memories should be stored only in a controlled self-hosted instance.
Redact memory URIs, node content, boot URIs, glossary triggers, search terms, database paths, API tokens, and snapshot diffs before sharing logs or screenshots.
Prerequisites
Python 3.10 or newer.
Node.js if the dashboard frontend needs to be built on first startup.
MCP client that supports stdio, SSE, or streamable HTTP depending on deployment mode.
SQLite or PostgreSQL storage selected and backed up before storing important memories.
API token configured before exposing the web, SSE, or HTTP server beyond localhost.
Schema details
Install type
cli
Troubleshooting
No
Source repository stats
Scope
Source repo
Collection metadata
Estimated setup
20 minutes
Difficulty
intermediate
Tool listing metadata
Disclosure
MIT-licensed open-source memory server. The project uses strong AI-persona framing in its README; this entry describes the source-backed MCP memory behavior without endorsing storing sensitive personal data unreviewed.
Nocturne Memory MCP Server is a self-hosted long-term memory server for agents.
It stores URI-addressed memory nodes, supports aliases and glossary triggers,
offers search and system views, and includes review and rollback mechanics for
changes made to the memory graph.
Use it when a team wants Claude or another MCP client to share durable memory
across sessions while keeping the memory database, dashboard, and write-review
process under local control.
These sources were reviewed on 2026-06-06. Prefer the live repository,
English README, license, MCP tool reference, MCP server implementation,
configuration, auth, SSE runner, snapshot store, and Python requirements for
current setup and behavior.
Features
Read a memory by URI with read_memory.
Create child memories with priority, disclosure text, and optional titles.
Update memory content through append or exact patch modes.
Delete a memory path without necessarily deleting the underlying node body.
Add alias paths that point to existing memories.
Manage glossary triggers for cross-linking memory recall.
Search memories by keyword, domain, and result limit.
Read system views such as boot memory, recent memories, glossary, indexes, and
diagnostics.
Use namespace isolation for separate agents or personas.
Review changes and roll back memory edits through snapshot tracking.
Run over stdio locally or expose SSE/streamable HTTP through the combined
server when properly authenticated.
Installation
Clone the repository, install backend dependencies, and configure an MCP client
to run the Python MCP server:
git clone https://github.com/Dataojitori/nocturne_memory.git
cd nocturne_memory
pip install -r backend/requirements.txt
For network-facing deployments, configure api_token, keep the host local unless
the server is intentionally exposed, and use the SSE or streamable HTTP runner
only behind reviewed access controls.
Use Cases
Give Claude durable project memory across coding sessions.
Store boot memories that orient an agent at the start of a session.
Keep separate memory namespaces for different agents, clients, or workspaces.
Search remembered facts without embedding the entire memory database in a
prompt.
Track memory changes before accepting or rolling them back.
Use glossary triggers to cross-link concepts that should be recalled together.
Run a visual dashboard for reviewing and editing the memory graph.
Safety and Privacy
Long-term memory changes future agent behavior. Read existing memories before
editing them, keep write operations small, and review snapshots before accepting
large restructures, cleanup, or deletion. Require an API token before exposing
Nocturne over SSE or HTTP, and keep local deployments bound to localhost unless
there is a clear access-control plan.
Treat the memory database as sensitive. Memory nodes may contain personal
history, private project strategy, credentials pasted by mistake, emotional
support context, relationship content, health details, drafts, and agent
instructions. Protect database files, PostgreSQL URLs, config files, API tokens,
snapshots, dashboard access, MCP transcripts, search results, and boot URIs.
Duplicate Check
No Dataojitori/nocturne_memory, Nocturne Memory MCP, rollbackable memory MCP,
agent memory graph, or matching source URL entry was found in content/mcp or
README.md. Existing memory entries cover different memory servers and do not
cover Nocturne's URI-addressed graph, dashboard, namespace, glossary, and
rollback workflow.
Show that Nocturne Memory MCP Server is listed on HeyClaude. Paste this Markdown into your README — it renders the badge and links back to this page.
[](https://heyclau.de/entry/mcp/nocturne-memory-mcp-server)
How it compares
Nocturne Memory MCP Server side by side with 3 alternatives on trust, install, platform support, and disclosed safety notes — all from reviewed registry metadata.
Long-term memory MCP server for agents, with URI-addressed memories, namespaces, search, aliases, glossary triggers, a visual dashboard, and reviewable rollback snapshots backed by SQLite or PostgreSQL.
Connect Claude to a Google Colab runtime. Colab MCP starts a local FastMCP server and websocket proxy, waits for an authorized Colab-origin browser connection, then exposes the session's notebook and runtime tools to your MCP client over stdio.
Django extension that exposes MCP endpoints and stdio transport for Django apps, with declarative model query tools, custom toolsets, DRF create/list/ update/delete tool publishing, serializer output, and MCP inspection.
MCP server example from FunASR that lets Claude transcribe local audio files with local speech recognition, automatic language handling, timestamps, and speaker labels when available.
✓Nocturne Memory can create, update, delete, alias, trigger, and search persistent memory nodes.
Stdio startup can launch a local admin dashboard, initialize the database, and build the frontend if dependencies are present.
Network-facing SSE or HTTP mode refuses to start without an API token when bound to a reachable host, but localhost-only deployments should still set one.
Write tools can change long-term agent behavior; review snapshots and rollback queues before accepting broad edits or cleanup operations.
Keep `public_readonly_mcp` enabled only for demos or intentionally read-only deployments.
✓Colab MCP runs a local websocket proxy and waits for a Google Colab browser session to connect before proxying session tools.
The server creates a bearer-style proxy token and accepts a single authorized Colab-origin websocket connection at a time.
A connected Colab runtime can execute code and interact with notebooks, outputs, files, variables, packages, and mounted resources depending on the session and tools exposed by Colab.
Use disposable notebooks or reviewed development sessions before allowing Claude to run code or inspect outputs.
Avoid connecting notebooks that have access to production data, cloud credentials, private datasets, paid accelerators, or long-running jobs unless the workflow has explicit approval.
✓Django MCP Server can expose Django model querysets, custom Python methods, DRF create/list/update/delete views, serializers, resources, and low-level FastMCP tools to an MCP client.
Published DRF create, update, and delete tools can mutate application data if their serializers, views, and authentication rules permit it.
The README notes that built-in DRF authentication classes, permission classes, filter backends, and pagination are disabled for published DRF tools in favor of MCP authentication; review this carefully before reusing production views.
Query tools can evaluate QuerySets and return database records; restrict queryset scope and fields before exposing sensitive models.
Require confirmation and application-level authorization before exposing write tools, email-sending methods, admin-like actions, or tools that touch customer, employee, financial, health, or regulated data.
✓The MCP server exposes a `transcribe_audio` tool that reads the local file path supplied by the agent.
Configure clients so Claude can only request audio files from approved directories; do not expose arbitrary private folders or shared drives.
First use can download FunASR model weights and dependencies from upstream model hosts; review network policy, cache location, and disk usage before use in restricted environments.
Long recordings and GPU transcription can consume significant CPU, GPU, memory, and disk cache resources.
Require confirmation before transcribing meetings, calls, interviews, voice notes, customer audio, regulated recordings, or files containing other people.
Privacy notes
✓Memories can contain personal history, health details, credentials accidentally pasted into chats, private project context, relationship content, strategies, drafts, and agent identity instructions.
Database files, PostgreSQL URLs, API tokens, config files, snapshots, dashboard views, MCP transcripts, and search results can expose the full memory graph.
The README includes public demo guidance, but private memories should be stored only in a controlled self-hosted instance.
Redact memory URIs, node content, boot URIs, glossary triggers, search terms, database paths, API tokens, and snapshot diffs before sharing logs or screenshots.
✓Notebook code, prompts, outputs, logs, variables, datasets, file paths, runtime metadata, package lists, browser session details, and generated websocket tokens can be visible to the MCP client and model provider.
Colab notebooks may contain API keys, OAuth tokens, mounted Drive paths, secrets in environment variables, private model weights, customer data, or unpublished research.
Local log files are created under a temporary Colab MCP log directory by default unless a log directory is specified.
Review Google Colab, MCP client, model provider, and organization retention policies before sending notebook or runtime context to an assistant.
✓Django sessions, request headers, model names, field names, primary keys, QuerySet results, serializer output, DRF request bodies, custom tool arguments, and tool responses can be exposed to the MCP client.
Exposed models may contain user accounts, permissions, customer records, orders, messages, files, logs, internal notes, audit trails, or application-specific secrets.
Remote streamable HTTP deployments can move application data outside the original Django UI and audit path if MCP auth, OAuth metadata, and retention are not configured correctly.
Stdio usage can still expose data through local MCP client logs, transcripts, and tool traces.
Keep MCP endpoint access, serializer fields, queryset filters, and tool docstrings intentionally narrow for each app.
✓Audio recordings can contain voices, names, accents, speaker identity, background speech, locations, health details, financial details, customer data, credentials spoken aloud, or other sensitive personal information.
The upstream MCP example performs local inference and does not require an API key, but MCP clients, model providers, logs, terminal output, transcripts, screenshots, and shared chats can still retain audio paths and transcription text.
Generated transcripts, timestamps, and speaker labels may identify individuals or reveal confidential conversations.
Model downloads and package installation can contact PyPI, ModelScope, Hugging Face, or other dependency hosts depending on the environment and model configuration.
Prerequisites
Python 3.10 or newer.
Node.js if the dashboard frontend needs to be built on first startup.
MCP client that supports stdio, SSE, or streamable HTTP depending on deployment mode.
SQLite or PostgreSQL storage selected and backed up before storing important memories.
uv or uvx available to run the server from the GitHub repository.
Local MCP client that supports `notifications/tools/list_changed`.
Google Colab account and browser session you are authorized to use.
Local machine access, because upstream documents that the MCP client must run locally.
Django 4 or 5 application with Python 3.10 or newer.
mcp_server added to INSTALLED_APPS and mcp_server.urls included in the Django URL configuration.
Review of which Django models, querysets, custom methods, DRF views, serializers, and request context should be exposed to MCP clients.
Authentication classes configured through DJANGO_MCP_AUTHENTICATION_CLASSES before exposing non-public data over streamable HTTP.
Python environment with FunASR installed from PyPI or a reviewed source checkout.
Local checkout or copy of `examples/mcp_server/funasr_mcp.py` from the FunASR repository.
Audio files in an approved location and format such as WAV, MP3, FLAC, M4A, or OGG.
Optional GPU, Apple silicon, or CPU device selection through `FUNASR_DEVICE`.
Install
Clone the repository, install `backend/requirements.txt`, and configure an MCP client to run `backend/mcp_server.py` with Python.