memini-ai-dev 1.3.0

Local-first semantic memory server with embedded PostgreSQL/pgvector. Optional RRF fusion with shared team server.

memini-ai-dev

"I remember" in Latin (pronounced meh-mee-nee)

Local-first semantic memory for AI agents. Vector search, trust scoring, knowledge graph, and persistent reasoning — fully MCP-compatible. Runs on embedded PostgreSQL (no Docker) or an external Postgres + pgvector server.

What it does

• Remembers facts, decisions, and patterns across sessions with trust-weighted retrieval
• Answers semantic queries with hybrid vector + BM25 search fused via RRF
• Tracks relationships between memories (SUPERSEDES, CONTRADICTS, DERIVED_FROM, RELATED_TO, PARTIAL_UPDATE)
• Detects and resolves contradictions via dialectic reasoning
• Generates project summaries (L0 ~100 tokens, L1 ~2K tokens) for session start
• Indexes codebases for semantic search across files
• Shares memories across agent peers with permission controls
• Decays low-trust memories and consolidates near-duplicates to keep the knowledge base relevant

Architecture

flowchart LR
    U[User] -->|prompt| OC[OpenCode TUI]
    OC -->|task| ORCH["Neuralgentics Orchestrator<br/>12 personas + routing matrix"]
    ORCH -->|query / save| MEM[("memini-ai<br/>FIRST-CLASS MCP<br/>registered directly in opencode.json")]
    MEM --> PG[("PostgreSQL + pgvector<br/>trust-weighted memory")]
    ORCH -->|dispatch| AG["Specialist sub-agents<br/>coder · architect · tester · writer"]
    AG -->|"long-tail tool calls"| BRK["Neuralgentics Broker<br/>catalog · access control · audit"]
    BRK --> MCP["Brokered MCP servers<br/>searxng · github · videre · ssh<br/>behind the broker · on demand"]
    AG -->|outbound HTTP| GW["Neuralgentics Gateway<br/>egress policy + audit"]
    GW --> NET["Internet / LLM APIs"]
    MEM --> WEB["Neuralgentics Web<br/>dashboards"]
    GW --> WEB
    BRK --> WEB

memini-ai is a first-class MCP server — registered directly in opencode.json and always loaded. Every other MCP server sits behind the broker: catalog-advertised, access-controlled, and brokered on demand, which keeps long-tail tool schemas out of every prompt.

See the memory lifecycle diagram for details.

Quickstart

Zero-setup (embedded PostgreSQL)

uvx --from memini-ai-dev memini-ai --stdio

On v1.x the embedded pgembed backend self-bootstraps on first run — no Docker, no external Postgres. Data lives under ~/.local/share/memini-ai/pgembed/data (XDG-compliant).

External PostgreSQL

export MEMINI_DB_URL="postgresql://user:password@localhost:5432/memini"
export MEMINI_VECTOR_BACKEND="postgres-external"

uvx --from memini-ai-dev memini-ai --stdio

v1.0.0 breaking change: if MEMINI_DB_URL is set you MUST also set MEMINI_VECTOR_BACKEND=postgres-external, otherwise the server raises RuntimeError: memini-ai v1.0.0: MEMINI_DB_URL is set but MEMINI_VECTOR_BACKEND is not. See CHANGELOG.md for the migration recipe.

Minimal MCP client config (opencode.json)

{
  "mcp": {
    "servers": {
      "memini-ai-dev": {
        "type": "local",
        "enabled": true,
        "environment": {
          "MEMINI_VECTOR_BACKEND": "pgembed",
          "TRUST_ENGINE": "true",
          "KG_ENABLED": "true"
        }
      }
    }
  }
}

Features

Memory Core

• Embeddings: MiniLM-L6-v2 (384-dim, default) or BGE-M3 (1024-dim, optional GPU upgrade). New deployments auto-upgrade to BGE-M3 when empty.
• Search: Hybrid vector + BM25 with Reciprocal Rank Fusion (RRF, k=60)
• Multi-model RRF: 384 + 1024 fusion in auto mode; optional 3rd CLIP image-recall arm when MEMINI_IMAGE_SEARCH_ENABLED=true
• Project isolation: Strict memory separation by MEMINI_PROJECT_ID

Trust Engine

Trust starts at 0.5 and is adjusted via feedback signals:

Signal	Delta
agent_used	+0.05
user_confirmed	+0.10
agent_ignored	-0.02
user_corrected	-0.15

Archive threshold: < 0.2. Promote to L1: > 0.8. Temporal decay fades irrelevant memories on a configurable half-life.

Memory Graph

Relationships: SUPERSEDES, PARTIAL_UPDATE, RELATED_TO, CONTRADICTS, DERIVED_FROM. Full supersession chain traversal (including archived memories). Live D3.js force-directed visualization.

Tiered Loading

• L0 Summary (~100 tokens): high-trust memories only (trust >= 0.5)
• L1 Key Decisions (~2K tokens): promoted memories (trust >= 0.8)
• L2 Full Context: all memories

Knowledge Graph

Entity extraction from memory text, inference chains between entities, formal KG queries with transitive closure, and HTML/D3.js graph export.

Thought Chains

Persistent multi-step reasoning with branching, revisions, and abandonment. API-compatible with sequential-thinking MCP servers.

Dialectic

Contradiction detection across memories, dialectic resolution synthesis, and per-memory argument history.

Decay and Consolidation

Temporal trust decay keeps the knowledge base relevant; consolidation merges near-duplicate memories on a configurable schedule.

Multi-Peer Sharing

Peer registration with trust levels, per-memory sharing with permission levels (shared, inherited, private), and opt-in per-project RBAC lockdown via MEMINI_PEER_ENFORCEMENT + MEMINI_PEER_ID.

MCP Tools (52)

See docs/mcp-tools.md for the full table. Categories: Basic Memory, Project Indexing, Trust and Tiering, Thought Chains, Knowledge Graph, Multi-Peer, User Modeling, System, and Advanced.

Storage

• Embedded PostgreSQL (default, no Docker): pgembed (Postgres 17, in-process, multi-process shared)
• External PostgreSQL: any PostgreSQL 16+ with pgvector and vectorscale — set MEMINI_VECTOR_BACKEND=postgres-external
• Optional team server RRF fusion: MEMINI_TEAM_DB_URL + MEMINI_FUSION_MODE=rrf

LLM Providers

• Local Ollama (default): any Ollama-compatible model
• Ollama Cloud: managed Ollama endpoints
• OpenAI-compatible: vLLM, LM Studio, and any OpenAI-compatible API

Configuration

Core

Variable	Default	Description
MEMINI_VECTOR_BACKEND	pgembed	pgembed (embedded) or postgres-external
MEMINI_DB_URL	(unset)	PostgreSQL URL (external mode only)
MEMINI_PGEMBED_DATA_DIR	~/.local/share/memini-ai/pgembed/data	Embedded Postgres data dir
MEMINI_TEAM_DB_URL	(unset)	Optional team server URL for RRF fusion
MEMINI_FUSION_MODE	none	none or rrf (fuses embedded + team)
MEMINI_EMBEDDING_DIM	384	384 or 1024
MEMINI_MODEL_NAME	all-MiniLM-L6-v2	HF model ID or alias (bge-m3, minilm)
MEMINI_EMBEDDING_MODE	auto	cpu, auto (384 + 1024 RRF), gpu
MEMINI_ENABLE_RRF	false	Enable RRF fusion across MiniLM + BGE-M3
MEMINI_PEER_ENFORCEMENT	false	Opt-in per-project RBAC lockdown
MEMINI_PEER_ID	(unset)	Project ID for RBAC tagging/filtering
LLM_PROVIDER	ollama	ollama, openai, or OpenAI-compatible
LLM_MODEL	llama3.2	Model name passed to the LLM provider
LLM_API_KEY	(unset)	API key (optional for local Ollama)
LLM_BASE_URL	(unset)	Override base URL for OpenAI-compatible

Feature Toggles

Feature	Env Var	Description
Trust Engine	TRUST_ENGINE	Trust scoring and archive/promotion
Tiered Loading	TIERED_LOADING	L0/L1/L2 summary generation
Knowledge Graph	KG_ENABLED	Entity extraction and KG queries
Memory Graph	MEMORY_GRAPH	Visual relationship mapping
Dialectic	DIALECTIC_ENABLED	Contradiction detection and resolution
Multi-Peer	MULTI_PEER_ENABLED	Peer-to-peer memory sharing
User Modeling	USER_MODELING	Persistent user profile tracking
Memory Decay	DECAY_ENABLED	Temporal trust decay engine
Auto-Extract	AUTO_EXTRACT	Auto extraction from conversations
Thought Chains	THOUGHT_CHAINS	Reasoning with branching/revision
Image Search	MEMINI_IMAGE_SEARCH_ENABLED	CLIP image-recall RRF arm

See the Configuration reference for the full list.

Documentation

• Architecture
• Configuration
• MCP Tools
• Upgrading Embeddings
• CHANGELOG

Development

uv sync          # install with dev dependencies
pytest           # run the test suite
ruff check src/  # lint
mypy src/        # type check

License

MIT License — see LICENSE for details.