Hindsight Mempalace

Name: Hindsight Mempalace
Author: holetron

@holetron

About Hindsight Mempalace

Self-hosted long-term memory for AI agents. An MCP server that gives agents persistent, hierarchical recall (L0–L3) over pgvector — retain preferences, decisions and observations, then recall them by topic across sessions. MIT, clone-and-run, requires Postgres + pgvector.

#memory #long-term-memory #pgvector #self-hosted #knowledge #rag #agents

Basic information

Config

Add this server to your MCP-compatible client using the configuration below.

{
  "mcpServers": {
    "hindsight-mempalace": {
      "command": "npx",
      "args": [
        "-y",
        "hindsight-mempalace-mcp"
      ],
      "env": {
        "HINDSIGHT_URL": "http://127.0.0.1:5100",
        "MEMPALACE_BANK": "mempalace-main"
      }
    }
  }
}

Tools

Save a memory with automatic room/hall classification

Scoped semantic search with room/hall/layer filters

Deep reasoning — synthesize facts, find patterns, answer with citations

Create closet summaries from accumulated facts

Cross-bank tunnels between related memories

Overview

What is Hindsight Mempalace?

Hindsight Mempalace is a hierarchical memory system for AI agents, merging Hindsight's long-term vector store with MemPalace's taxonomy architecture. It provides structured memory with semantic search, rooms, halls, layers, closets, and tunnels, plus an MCP server for client integration.

How to use Hindsight Mempalace?

Clone the repository, copy .env.example to .env, edit configuration, then run docker compose -f docker-compose.mempalace.yml up -d. API available at http://localhost:5100. For the MCP server, install npm dependencies in mcp-server/ and run with environment variables.

Key features of Hindsight Mempalace

Vector store with hierarchical taxonomy (rooms, halls, layers)
Keyword-based auto-classification (<1ms, zero LLM cost)
L0–L3 priority tiers (always loaded to archive)
Cross-bank tunnels for multi-agent memory sharing
Closet compression: summaries with source pointers
5 MCP tools: retain, recall, reflect, compress, bridge

Use cases of Hindsight Mempalace

AI agents needing structured long-term memory with topic isolation
Multi-agent systems sharing insights across banks via tunnels
Development teams storing deployment procedures, warnings, and decisions
Knowledge management with automatic classification and compression

FAQ from Hindsight Mempalace

What does Hindsight Mempalace add over vanilla Hindsight?

It adds a taxonomy layer (rooms, halls, layers), auto-classification, closets, tunnels, cross-bank bridges, and a standalone MCP server with 5 tools.

What are the runtime requirements?

Docker for the API server; Node.js for the MCP server. Ships with BAAI/bge-small-en-v1.5 embeddings (English-optimized). Optional multilingual model BAAI/bge-m3 for other languages.

How does data classification work?

A keyword-based classifier (room_hall_classifier.py) assigns room and hall automatically in under 1ms without LLM calls. Users can also specify these manually.

Can I use this as a drop-in replacement for Hindsight?

Yes. The base /retain and /recall endpoints are fully backward-compatible. All new parameters are optional.

How do tunnels and bridges work?

Tunnels are cross-bank bridges between agents. Agent A discovers a memory; Agent B accesses it through a tunnel without duplicating data. Created via the /bridge endpoint or memory_bridge MCP tool.

What models are supported for embeddings?

Default English model BAAI/bge-small-en-v1.5 (384-dim). For multilingual use, set HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL=BAAI/bge-m3 (1024-dim) on an empty store.