Local AI Memory System

Your AI
Remembers
Everything

Full conversation memory, local-first privacy, and instant findability. The reasoning, decisions, and debugging history from your AI sessions — preserved exactly as they happened.

Install MemPalace View Benchmarks
0%
LongMemEval R@5
Raw Mode
0
Questions
Tested
Zero
API Calls
Required
Free
No Cloud
No Subscription
The Problem

When the Session Ends,
the Reasoning Dies

Important decisions now happen inside AI conversations. Architecture debates. Tradeoff analysis. Failed attempts you learned from. When those sessions end, all of it disappears — reduced to thin summaries that lose the context that actually mattered.

Other memory products summarize your history into something that sounds useful but isn't. The original reasoning, the dead ends, the moment the decision actually got made — gone.

Star field — vast, dark, and deep
Core Solution

A Structured Memory
Palace for Your AI

MemPalace keeps the original exchanges instead of throwing them away, then makes them findable through a structured memory model. Not summaries. The actual source.

Layer 01

Wings

Top-level containers scoped to people and projects. Every wing holds all memory associated with a single context — a team, a codebase, a client.

Layer 02

Rooms

Specific ideas and topics live inside rooms. A room might hold everything about a particular architecture decision or a recurring debugging problem.

Layer 03

Halls

Memory organized by type — factual records, episodic conversations, procedural patterns. Halls are how the system knows where to look when you ask a question.

Layer 04

Drawers

The verbatim source. Every drawer preserves the exact conversation or file that a fact came from — retrievable, linkable, and unchanged.

Layer 05

Tunnels

Connections between rooms and wings. When a decision in one project affected another, tunnels make those links explicit and traversable.

Layer 06

Closets

Lightweight temporary storage for context that hasn't been categorized yet. Auto-save hooks populate closets during active work sessions.

How It Works

Three Moves.
Complete Memory.

01

Wake Up

Your AI loads a lightweight context layer — identity and critical facts — so it always knows who you are and what matters without reading everything.

02

Search Deep

When a question needs history, MemPalace searches targeted wings and rooms on demand. Fast, specific, and only as deep as needed.

03

Recover Exact

The system surfaces the original conversation or file behind a decision — not a summary of it. The full verbatim source, linked and ready.

Your Workflow

Two Paths.
One Memory System.

Path 01 — MCP-Compatible

AI Tools

Connect MemPalace through MCP-compatible tools. Your AI uses memory automatically during sessions — no manual commands needed.

Claude ChatGPT Cursor Gemini
Path 02 — Offline Models

Local CLI

For offline and local models, use wake-up context injection and targeted search via CLI or Python API. Full capability with zero cloud dependency.

Python API CLI Offline
Satellite orbiting Earth — persistent connection across distance
Feature Layer

Built Around
the Core Engine

Temporal Knowledge Graph

Facts don't just get stored — they get timestamped and tracked over time. When a decision changes, the graph knows when it changed and why. Query the state of any fact at any point in your project history.

Agent Diaries

Specialist agents keep memory inside their own domain. Each agent maintains its own history without polluting the shared context.

Auto-Save Hooks

Memory builds during work without manual effort. Hooks capture useful context automatically so your palace grows as you build.

Real-World Retrieval

The Value Is in
the Recovery

01
Solo Builder

Why Did We Choose Postgres?

Six months later, a new teammate asks why the team chose Postgres over MongoDB. MemPalace surfaces the original conversation — tradeoffs, load estimates, the moment the decision landed.

02
Team Context

What Did Alex Work On Last Sprint?

Pull the exact sessions, files, and decisions from a specific person's work in a specific period. Not a summary. The actual history, filtered and ready.

03
Cross-Project

Who Changed the Auth Flow?

Trace a change across projects and teams. MemPalace follows tunnels between wings to surface the decision chain — even when it spans months and multiple codebases.

Benchmarks
96.6%
LongMemEval R@5 — Raw Mode — 500 Questions

Reproducible results. Zero API calls. Tested on the standard LongMemEval benchmark suite with no proprietary data or external lookups. Run it yourself.

View Reproducibility Guide
Raw Mode R@5 96.6%
Questions Tested 500
External API Calls 0
Rerank Mode R@5

Requires additional API calls. Different mode — not the default product claim.

100%
Transparency

We Say What
We Don't Know

MemPalace ships with an honest changelog. Not because we have to — because the alternative is a product you can't trust.

What We've Corrected
Updated

AAAK (Adaptive Associative Anchor Keys) was overstated in the original README. It is lossy rather than lossless and can regress benchmark performance in some configurations. The 34% structure boost claim was framed too aggressively and lacked adequate qualification.

Contradiction detection is not yet wired into the main knowledge graph flow. This is a roadmap item, not a current feature. We've removed it from the primary feature list until it ships.

The 100% rerank result requires additional API calls and is a separate mode from the default product. We now label it clearly rather than mixing it into the default claim.

What Remains True

96.6% LongMemEval R@5 in raw mode is a stable, reproducible result. The benchmark is standard, the test set is 500 questions, and the methodology is documented. Zero external API calls. Run it yourself.

The memory model — wings, rooms, halls, drawers, tunnels, closets — is fully implemented and tested. The verbatim retrieval system works as described. Local-first architecture with no cloud dependency is the foundation, not a marketing claim.

Open Source · Local-First · Yours

Install MemPalace.
Keep Every Conversation.

Runs on your machine. No subscription. No cloud. Connect to your AI workflow in minutes and start building memory that doesn't disappear.

Get Started — Install Now Read the Docs
This report was created by Xavior
AI with the tools to actually act.
Sends emails. Manages your calendar. Writes the next one.
Try free → xavior.ai