Installs into .claude/skills of the current project.
Are you the author of Server Configuration And Memory Engines?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/vectorspacelab-server-configuration-and-memory-engines)
---
name: server-configuration-and-memory-engines
description: "Use the MemMachine server, configuration, REST API, MCP entry
points, memory engines, storage/provider resources, Docker/Helm deployment
shape, and backend troubleshooting."
disable-model-invocation: true
metadata:
disco-role: operating
license: Apache 2.0
---
# Server Configuration And Memory Engines
Use this sub-skill when a task names `memmachine-server`, MemMachine REST API,
MCP stdio/HTTP, `cfg.yml`, Docker Compose, Helm, storage resources, provider
resources, episodic memory, semantic memory, short-term memory, retrieval-agent
mode, or server startup/troubleshooting.
## Route Within This Sub-skill
- Read [server-configuration.md](references/server-configuration.md) for
configuration sections, startup modes, environment variables, and safe
validation.
- Read [storage-and-providers.md](references/storage-and-providers.md) for
database/vector/graph stores, embedders, language models, rerankers, optional
extras, and credential boundaries.
- Read [rest-and-mcp.md](references/rest-and-mcp.md) for REST endpoint families,
health/metrics, MCP tools, and context parameters.
- Read [memory-engines.md](references/memory-engines.md) for event vs
declarative long-term memory, short-term memory, semantic/profile memory, and
retrieval-agent behavior.
- Read [troubleshooting.md](references/troubleshooting.md) for server startup,
config, Docker, storage, provider, MCP, and optional dependency failures.
- Run [server_config_doctor.py](scripts/server_config_doctor.py) to validate a
MemMachine YAML config shape without starting services.
- Run [mcp_entrypoint_smoke.py](scripts/mcp_entrypoint_smoke.py) to inspect MCP
entry-point availability and context parameters without launching a listener.
## Safe Startup Checks
A local server task usually starts with read-only checks:
```bash
memmachine-server --help
memmachine-mcp-http --help
python scripts/server_config_doctor.py --config cfg.yml
python scripts/mcp_entrypoint_smoke.py --check-imports
```
Only start or stop services after the user confirms the desired deployment and
side effects. Docker Compose and Helm workflows can create volumes, containers,
network listeners, and persistent data.
## Configuration Shape
A typical server config has these top-level sections:
```yaml
logging: {}
episode_store: {}
episodic_memory: {}
retrieval_agent: {}
semantic_memory: {}
session_manager: {}
resources:
databases: {}
embedders: {}
language_models: {}
rerankers: {}
```
Important backend decision:
- `episodic_memory.long_term_memory.backend: declarative` uses a
`vector_graph_store` resource such as Neo4j/Nebula.
- `backend: event` uses a `vector_store` plus a `segment_store` resource.
## Server Commands And Entry Points
- `memmachine-server --help` shows server flags including `--stdio`,
`--version`, and `--with-config-api`.
- `memmachine-mcp-stdio` runs MCP in stdio mode for local MCP clients.
- `memmachine-mcp-http --host HOST --port PORT` runs MCP over HTTP.
- `memmachine-configure` is interactive; do not run it in automation unless the
user explicitly wants an interactive installer/configurer.
## Cross-links
- Use [python-sdk-and-cli](../python-sdk-and-cli/SKILL.md) for application-side
Python client calls, `mem-cli`, and LangGraph helper usage.
- Use [integrations-and-migration](../integrations-and-migration/SKILL.md) for
framework/platform recipes and conversation import tools.
- Use [typescript-rest-client](../typescript-rest-client/SKILL.md) for Node/TS
client calls against a MemMachine server.
- Use the root [REST API and data models](../../references/rest-api-and-data-models.md)
reference when translating endpoint payloads across clients.
## Verification Boundaries
- Import and help checks prove packages and entry points are available; they do
not prove databases, providers, or MCP clients are configured.
- Docker, Neo4j, Postgres, Qdrant, Milvus, Nebula, OpenAI, AWS Bedrock, Cohere,
and Ollama/OpenAI-compatible endpoints are optional external dependencies.
- The optional spaCy multi-hop decomposer may be absent in base installs. Do not
classify its warning as a fatal server failure unless the user specifically
needs spaCy-based decomposition.