Configuration¶

memini-ai is configured via environment variables (or a JSON config file at MEMINI_CONFIG_PATH). All settings are optional - defaults work for a solo local-first install.

Core settings¶

Tool surface gating (v1.7.0+)¶

memini-ai registers 52 MCP tools by default. You can restrict the exposed surface so your MCP client sends fewer schemas and tokens per request. There are two mechanisms; the more specific one wins.

1. Per-tool allow-list — MEMINI_ENABLED_TOOLS¶

Set to a comma-separated list of exact tool names. Only those tools are registered. This overrides MEMINI_TOOL_GROUPS.

export MEMINI_ENABLED_TOOLS="query_memories,add_memory"

In opencode.json:

{
  "mcp": {
    "servers": {
      "memini-ai-dev": {
        "type": "local",
        "enabled": true,
        "environment": {
          "MEMINI_ENABLED_TOOLS": "query_memories,add_memory"
        }
      }
    }
  }
}

• Names are the exact function __name__ values used at registration time (e.g. query_memories, add_memory, search_project).
• Whitespace around commas is trimmed.
• Unknown names are ignored with a warning log.
• If unset or empty, the server falls back to MEMINI_TOOL_GROUPS.

2. Group-based gating — MEMINI_TOOL_GROUPS¶

Coarse groups exposed when MEMINI_ENABLED_TOOLS is unset.

Default: core,trust,kanban,session.

export MEMINI_TOOL_GROUPS="core,trust,chains,kg"

Unknown group names are ignored with a warning log. | MEMINI_CHUNK_OVERLAP | 50 | Overlap between chunks. | | MEMINI_BATCH_SIZE | 32 | Batch size for embedding generation. | | MEMINI_WORKERS | cpu_count | Number of worker threads. | | MEMINI_LOG_LEVEL | info | Logging level (debug, info, warning, error). | | MEMINI_DEVICE | auto | Device for embeddings (auto, cpu, cuda). | | MEMINI_CONFIG_PATH | (empty) | Path to a JSON config file. |

Feature toggles¶

All features below are disabled by default. Set to true to enable.

Trust engine tuning¶

Image recall (v0.8.0+)¶

TLS / SSL¶

PostgreSQL connections support TLS encryption to prevent data exfiltration.

SSL mode values (from the libpq docs):

RBAC: per-project users¶

By default all users have read/write/edit access to all memories (open access). Per-project isolation is opt-in.

User management CLI:

memini-ai user add --name <name> [--admin]    # Add a user
memini-ai user list                            # List all users
memini-ai user remove --name <name>            # Drop a user

The .env file pattern¶

• Homedir .env (~/.config/opencode/.env): admin credentials.
• Project .env (./.env): project credentials.
• opencode.json references via {env:MEMINI_PROJECT_USER} etc. - secrets are never committed to git.
• .env files are chmod 600.

Quick start with all features on¶

export MEMINI_TRUST_ENGINE=true
export MEMINI_TIERED_LOADING=true
export MEMINI_KG_ENABLED=true
export MEMINI_DIALECTIC_ENABLED=true
export MEMINI_DECAY_ENABLED=true
export THOUGHT_CHAINS=true
export MEMINI_MULTI_PEER_ENABLED=true
uvx --from memini-ai-dev memini-ai --stdio