mnemos
📍 Canonical source: GitLab
The authoritative source for this project lives on GitLab — always: https://gitlab.com/ncz-os/mnemos
This GitHub repository is retained only to host container images on ghcr.io (
mnemos-core,mnemos,mnemos-enterprise). Development, issues, and merge requests happen on GitLab.
MNEMOS + GRAEAE
MNEMOS v6.1.7 is the memory operating system for serious agentic work. It is not just a place to put bytes: it is a runtime of named subsystems that manage the full lifecycle of agent memory across providers, agents, and time horizons — write, embed, search, compress, version, reason over, audit, federate, export, import, and operate.
What is in the box:
a packaged FastAPI runtime with a CLI-first deployment surface
EPIMONE, the six-backend persistence layer — SQLite + sqlite-vec by default, PostgreSQL + pgvector, Oracle Database 26ai, IBM Db2 12.1, MySQL 9.0 Enterprise/HeatWave, and MariaDB 11.7+. Every backend self-provisions its schema on first connect. See Persistence.
the GRAEAE reasoning bus and PANTHEON unified LLM facade
an operator-audited compression stack
a divergent dream-state pipeline: REPLAY → CLUSTER → CONSOLIDATE → SYNTHESISE → EXTRACT
a GDPR right-to-be-forgotten worker
the PERSEPHONE archival subsystem
KRONOS recall observability
How it is packaged. MNEMOS ships as a small core (
mnemos-core) plus separately installablemnemos.*namespace subsystems (GRAEAE, PANTHEON, KNEMON, CHARON) and the standalone STIPHOS hive service. Container images are published toghcr.io/ncz-os—mnemos-core,mnemos(amd64 + arm64), andmnemos-enterprise(amd64, adds the Oracle/Db2/MySQL drivers). Pin the minor line (:6.1) to keep a fleet on identical code. Install and DSN/driver setup for each backend are in docs/INSTALL.md; the agent-facing contract is in AGENTS.md.
Quick Start
🚀 Fastest path — the free-backend Quickstart. Two commands to durable, MCP-accessible agent memory on a free database. It's built around IBM Db2 12.1 — the reference deployment for the "mnemos on Db2" IBM TechXchange write-up — but the exact same image and steps run unchanged on Oracle Database 23ai Free, PostgreSQL + pgvector, or MariaDB 11.7+. Pick a backend in
quickstart/docs/BACKENDS.md; we lead with Db2, it works with all of them.
Memory and reasoning runtime for AI agents: persistent search, versioned storage, webhook fanout, and a unified LLM routing bus - all behind a single MCP interface.
1. Agent-driven install
Paste into Claude Code, Cursor, or Codex. The agent runs the install; you confirm. Agents should read AGENTS.md — it has the machine-readable module registry and a deterministic procedure for installing exactly the requested modules on the operator's arch + backend.
The pip package is mnemos-core (the subsystems are separate dists pulled
via extras). mnemos is the published image name, not a pip package.
Turnkey (container, any arch):
docker run -p 5002:5002 -v mnemos-data:/data ghcr.io/ncz-os/mnemos:latest
# everything image: core + graeae + pantheon + knemon + charon. SQLite by default.
# Point at a real DB with -e MNEMOS_DATABASE_DSN='postgres://…' (or oracle://… thin).pip (compose your own):
Install MNEMOS on this machine.
Steps:
1. pip install 'mnemos-core[server]' # everything; arch-neutral (no openvino)
2. mnemos init # scaffold config + token
3. mnemos serve # start API on :5002
4. mnemos doctor # verify subsystems
5. Set MNEMOS_BASE=http://localhost:5002 and MNEMOS_API_KEY=<token from step 2>
in shell env and any agent config that needs to reach it.
Edge device (SQLite kernel only): pip install 'mnemos-core[edge]'
Single subsystem, e.g. reasoning: pip install 'mnemos-core[graeae]'
Hive (STIPHOS) is a SEPARATE service: pip install 'mnemos-stiphos[mcp]' (port 8080)Enterprise backends (Oracle Database 26ai, IBM Db2 12.1).
Turnkey is the amd64-only mnemos-enterprise image (everything + Oracle/Db2/
MySQL drivers baked in). Note: Oracle uses the thin driver, so it also runs on
the plain mnemos image and on arm64 — only Db2 actually requires enterprise.
See docs/INSTALL.md for full driver, DSN, and migration
steps.
# Turnkey (amd64):
docker run --platform linux/amd64 -p 5002:5002 \
-e MNEMOS_DATABASE_DSN='db2://user:pass@host:50000/dbname' \
ghcr.io/ncz-os/mnemos-enterprise:latest
# Or from source:
git clone https://github.com/ncz-os/mnemos && cd mnemos
python -m pip install -e '.[server,enterprise]' # or '.[server,oracle]' / '.[server,db2]'
export MNEMOS_DATABASE_DSN='oracle://user:pass@host:1521/service_name'
# or: MNEMOS_DATABASE_DSN='db2://user:pass@host:50000/dbname'
mnemos install --profile server
mnemos serve --profile server2. Connect an agent via MCP
Add to ~/.claude/mcp_servers.json (Claude Code) or equivalent:
{
"mcpServers": {
"mnemos": {
"command": "mnemos",
"args": ["serve", "mcp-stdio"],
"env": {
"MNEMOS_BASE": "http://<host>:5002",
"MNEMOS_API_KEY": "<token>"
}
}
}
}For HTTP/SSE transport (ChatGPT, remote agents): mnemos serve mcp-http, which listens on :5003 by default.
Key MCP tools the agent gets:
Tool | What it does |
| Semantic + filtered search across the memory store |
| Write a new memory with category, tags, and content |
| Fetch a memory by ID |
| Query the knowledge-graph triple store |
| Surface recall anomalies and memory health signals |
| List soft-deleted memories pending hard deletion |
3. Webhooks + integrations
Integration | What connects | How |
Claude Code | Hooks fire on session-start, prompt-submit, stop - auto-log to MNEMOS |
|
ZeroClaw | Zeroclaw agent reads/writes memories via MCP |
|
OpenClaw | OpenClaw gateway routes memory ops through MCP |
|
Hermes | Optional memory skill mounts MNEMOS as a tool provider |
|
Webhooks (any) | Push |
|
Cursor / Cline / Continue.dev / Zed / Aider | Any MCP-capable IDE connects via stdio or HTTP transport | See |
Full documentation: docs/
Architecture
MNEMOS is a packaged FastAPI service with a single mnemos CLI for
installation, serving, MCP transport, and operational checks. Agents connect
through MCP stdio, MCP HTTP/SSE, REST, or OpenAI-compatible SDKs, while the
runtime routes memory, reasoning, session, webhook, federation, portability,
and observability work through the mnemos/ package. GRAEAE handles
multi-provider reasoning and model routing; MOIRAI handles operator-audited
compression through APOLLO and ARTEMIS.
Persistence
The backend is chosen at runtime by DSN scheme, never by rebuild. Six are
implemented, in mnemos/persistence/:
Backend | Vector support | Notes |
SQLite + sqlite-vec |
| Default. Edge and development installs; no server to run. |
PostgreSQL + pgvector | HNSW | Recommended for vector and semantic workloads — the most mature and predictable option, with broad managed-service support. |
Oracle Database 26ai | HNSW | Also JSON Duality and TDE. Thin driver, so it runs on the standard |
IBM Db2 | DiskANN | Hot paths emit native Db2 SQL — |
MySQL 9.0+ |
| For the managed-cloud MySQL audience (RDS and Aurora MySQL, HeatWave). Note that the vector functions ship only in MySQL Enterprise/HeatWave, not Community. |
MariaDB 11.7+ |
| The strongest MySQL-family option, and available in the free Community edition. Embeddings live in a |
Every backend satisfies the same PersistenceBackend protocol set
(mnemos/persistence/base.py) and self-provisions its schema idempotently on
backend.open(), DSN-aware, with the dimension taken from
MNEMOS_EMBEDDING_DIM. SQLite, PostgreSQL, Oracle, and Db2 share the
cross-backend harness in tests/test_persistence_parity.py; MySQL and MariaDB
are covered by their own live suites.
Documentation
Topic | File |
Installation | |
Specification | |
System requirements | |
Memory architecture | |
Compression | |
GRAEAE reasoning | |
PANTHEON provider facade | |
KRONOS observability | |
Audit chain | |
Portability format (MIF 1.0) | |
Scaling | |
Operations | |
Benchmark harness | scripts/bench_v4.py — cross-backend vector-search harness (PG / Oracle / Db2 / SQLite). Results published post-GA. |
License
MNEMOS is licensed under the Apache License, Version 2.0. See LICENSE for the full text.
Build infrastructure & partners
Continuous integration and package distribution for this project are generously supported by our open-source infrastructure partners:
GitLab — canonical source hosting and CI pipelines (format / lint / test gates), via the GitLab for Open Source program.
Buildkite — CI/CD orchestration with hosted macOS and Linux agents, and our APT package registry host (
packages.buildkite.com/ncz-os/ncz), via the Buildkite Open Source program.
Thank you to both for backing open-source software.
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/ncz-os/mnemos'
If you have feedback or need assistance with the MCP directory API, please join our Discord server