memini-ai¶
"I remember" in Latin (pronounced meh-mee-nee)
Local-first semantic memory server with vector search, trust scoring, and persistent reasoning, fully MCP-compatible.
Why memini-ai¶
Every long-running agent hits the same wall: the context window fills up, the earlier half of the conversation is evicted, and the agent starts repeating itself or forgetting decisions. memini-ai gives your agent a durable, queryable memory backed by PostgreSQL + pgvector, with a trust engine so good memories get reinforced and bad ones decay.
Key features¶
- MCP-compatible - drop-in for OpenCode, Claude Desktop, or any MCP client
- Trust engine - every memory starts at trust 0.5 and is adjusted by agent and user feedback (
agent_used+0.05,user_confirmed+0.10, etc.) - Tiered memory - L0 (~100 tokens, summary), L1 (~2K, key decisions), L2 (full context) for efficient context loading
- Knowledge graph - entity extraction, typed relationships, live D3.js visualization, inference chains between entities
- Dialectic reasoning - automatic contradiction detection and LLM-driven resolution between conflicting memories
- Memory decay - temporal trust decay keeps the store focused on what still matters
- Multi-peer - share memory subsets across distinct agent personas
- Dual-model RRF - 384-dim MiniLM (CPU default) plus optional 1024-dim BGE-M3 (GPU), fused with Reciprocal Rank Fusion (k=60)
- Persistent thought chains - branching, revisable reasoning chains that survive across sessions
- Project isolation - strict per-project memory separation via
project_id
Install¶
# Recommended: run via uvx (no install needed)
uvx --from memini-ai-dev memini-ai --stdio
# Or install with pip
pip install memini-ai-dev
memini-ai --stdio
memini-ai v1.x self-bootstraps an embedded PostgreSQL 17 + pgvector server on first run - no Docker, no external database required. The data directory lives at ~/.local/share/memini-ai/pgembed/data by default.
First MCP call¶
Point any MCP client at the memini-ai --stdio command and call:
{
"method": "tools/call",
"params": {
"name": "add_memory",
"arguments": { "content": "memini-ai is now installed" }
}
}
Then retrieve it:
{
"method": "tools/call",
"params": {
"name": "query_memories",
"arguments": { "query": "install status" }
}
}
Next steps¶
- Getting started - full install, bootstrap, and config
- Configuration - every env var and feature toggle
- MCP tools - all 52 tools grouped by category
- Architecture - components and memory lifecycle
- Upgrading embeddings - MiniLM to BGE-M3 migration
- Changelog - per-release history