Skip to main content
Glama

📍 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 the mnemos-enterprise container image on ghcr.io. Development, issues, and merge requests happen on GitLab.


MNEMOS + GRAEAE

MNEMOS v7.0.0 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 AI 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

  • first-party CHARON portability: MIF/MPF import and export, universal ingest, migrate-in adapters, and STYX encrypted off-fleet backups

  • KRONOS recall observability

How it is packaged. MNEMOS ships as a small core (mnemos-core) plus separately installable mnemos.* namespace subsystems (GRAEAE, PANTHEON, KNEMON) and the standalone STIPHOS hive service. CHARON's portability, migration, ingestion, and STYX backup code became first-party core modules on 2026-09-18; only Docling's heavy document-conversion dependencies remain optional through mnemos-core[docling]. The published container image is ghcr.io/ncz-os/mnemos-enterprise — a single multi-arch (amd64 + arm64) manifest with every backend driver (Oracle, MySQL, MariaDB) except Db2, which is amd64-only. Pin an exact version (:7.0.0) 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.

Related MCP server: SyncContext

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 AI Database 26ai 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-enterprise:latest
# everything image: core (including CHARON/STYX) + graeae + pantheon + knemon. SQLite by default.
# Point at a real DB with -e MNEMOS_DATABASE_DSN='postgres://…' (or oracle://… thin).

pip (compose your own):

None of mnemos-core, mnemos-graeae, mnemos-pantheon, mnemos-knemon, or mnemos-stiphos are currently published to PyPI. pip install 'mnemos-core[...]' will 404 — the [server]/[full] extras recurse into these names on the public index. Until they're published, install from source. This exact sequence is tested in a clean venv:

# Core (arch-neutral, no openvino):
git clone https://gitlab.com/ncz-os/mnemos && cd mnemos
python -m pip install -e .

# Add-ons, straight from GitLab — core's already-installed version satisfies
# each add-on's own mnemos-core floor, so pip resolves it locally and never
# needs to reach PyPI for mnemos-core itself:
pip install 'git+https://gitlab.com/ncz-os/graeae.git'
pip install 'git+https://gitlab.com/ncz-os/knemon.git'
pip install 'git+https://gitlab.com/ncz-os/pantheon.git'   # needs graeae + knemon installed first

mnemos init                         # scaffold config + token
mnemos serve                        # start API on :5002
mnemos doctor                       # verify subsystems
# Set MNEMOS_BASE=http://localhost:5002 and MNEMOS_API_KEY=<token from mnemos init>
# in shell env and any agent config that needs to reach it.

# Hive (STIPHOS) is a SEPARATE service, same pattern:
git clone https://gitlab.com/ncz-os/mnemos-stiphos && cd mnemos-stiphos
pip install -e '.[mcp]'   # port 8080

Run pip check afterward as a sanity check — a clean install reports "No broken requirements found." Anything it does report is a real gap; install it.

Edge device (SQLite kernel only): after the core install above, pip install aiosqlite sqlite-vec (the [edge] extra's own two deps — both real, published packages, unaffected by the above).

Enterprise backends (Oracle AI Database 26ai, IBM Db2 12.1). mnemos-enterprise is a single multi-arch (amd64 + arm64) image with every backend driver baked in. The one asymmetry: Db2's driver has no arm64 wheel, so the Db2 backend is amd64-only — Oracle (thin driver), MySQL, and MariaDB all work on arm64 too. 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 (see the "pip (compose your own)" section above for the
# add-on install order — enterprise adds the same set plus the driver extra
# on core itself):
git clone https://gitlab.com/ncz-os/mnemos && cd mnemos
python -m pip install -e '.[enterprise]'   # or '.[oracle]' / '.[db2]' — core's own extras, real PyPI deps
pip install 'git+https://gitlab.com/ncz-os/graeae.git'
pip install 'git+https://gitlab.com/ncz-os/knemon.git'
pip install 'git+https://gitlab.com/ncz-os/pantheon.git'
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 server

2. 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

search_memories

Semantic + filtered search across the memory store

create_memory

Write a new memory with category, tags, and content

get_memory

Fetch a memory by ID

kg_search

Query the knowledge-graph triple store

kronos_anomalies

Surface recall anomalies and memory health signals

list_deletions

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

integrations/claude-code/ - copy hooks + set MNEMOS_BASE

ZeroClaw

Zeroclaw agent reads/writes memories via MCP

integrations/zeroclaw/ + mnemos serve mcp-stdio in zeroclaw config

OpenClaw

OpenClaw gateway routes memory ops through MCP

integrations/openclaw/ + MCP server entry in openclaw.json

Hermes

Optional memory skill mounts MNEMOS as a tool provider

integrations/hermes/optional-skills/memory/mnemos/

Webhooks (any)

Push memory.created, memory.updated, memory.deleted, consultation.completed events to any HTTPS endpoint

POST /v1/webhooks with {"url": "...", "events": [...]}

Cursor / Cline / Continue.dev / Zed / Aider

Any MCP-capable IDE connects via stdio or HTTP transport

See docs/connectors/


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

vec0 virtual table

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 AI Database 26ai

HNSW INMEMORY NEIGHBOR GRAPH

Also JSON Duality and TDE. Thin driver, so it runs on the standard mnemos image and on arm64.

IBM Db2

DiskANN

Hot paths emit native Db2 SQL — VECTOR_DISTANCE(..., EUCLIDEAN) with FETCH APPROX FIRST, engaging the DiskANN index on the user-facing query path. The default dialect (MNEMOS_DB2_DIALECT=compat) still translates inherited Oracle-shaped SQL at the cursor layer; set MNEMOS_DB2_DIALECT=native for the pass-through backend. amd64 only.

MySQL 9.0+

VECTOR_DISTANCE

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+

VEC_DISTANCE_COSINE + HNSW VECTOR INDEX

The strongest MySQL-family option, and available in the free Community edition. Embeddings live in a memory_embeddings join table. Its vector engine is newer than pgvector's and correspondingly less battle-tested.

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 and PostgreSQL share the cross-backend harness in tests/test_persistence_parity.py; Oracle, Db2, MySQL, and MariaDB are each covered by their own live suite.

Documentation

Topic

File

Installation

docs/INSTALL.md

Specification

docs/SPECIFICATION.md

System requirements

SYSTEM_REQUIREMENTS.md

Memory architecture

docs/MEMORY_ARCHITECTURE.md

Compression

docs/COMPRESSION.md

GRAEAE reasoning

docs/GRAEAE_FEATURES.md

PANTHEON provider facade

docs/PANTHEON.md

KRONOS observability

docs/KRONOS.md

Audit chain

docs/AUDIT_CHAIN.md

Portability format (MIF 1.0)

docs/MEMORY_EXPORT_FORMAT.md

Scaling

docs/SCALING.md

Operations

docs/OPERATIONS.md

Backend parity matrix

docs/BACKEND_PARITY.md — generated by scripts/generate_backend_parity_matrix.py, CI-gated against drift by lint:backend-parity. Do not hand-edit.

Known limitations

KNOWN_LIMITATIONS.md

SQLite / edge profile

docs/SQLITE_PROFILE.md

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.

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    C
    maintenance
    A production-ready MCP server that enables multiple AI agents to collaborate through a shared, concurrency-safe memory space. It supports advanced search, full CRUD operations, and automatic backups to facilitate asynchronous communication between agents.
    MIT
  • A
    license
    Not graded
    quality
    C
    maintenance
    Universal MCP server providing adaptive semantic memory for AI agents, supporting document ingestion, semantic search, chat persistence, cross-project linking, and cloud storage overflow.
    3
    GPL 3.0
  • A
    license
    Not graded
    quality
    B
    maintenance
    Persistent memory MCP server for AI agents that stores, recalls, and searches conversation history, key-value context, and long-term entries across sessions with semantic search and FIFO queues.
    61 npm
    1
    -