Project Manifesto¶

Memory Daemon¶

A fully local cognitive memory architecture.

Goal¶

Provide an extensible long-term memory system for language models.

The memory system should remain independent of any individual LLM.

The LLM is replaceable.

Memory is not.

Philosophy¶

The system is built around one idea:

Reasoning should not depend on memory.

Memory should support reasoning.

Never own it.

What This Means¶

Reasoning should not depend on memory. - You should be able to reason without retrieving - Reasoning is a separate cognitive function - Memory is a substrate, not a crutch

Memory should support reasoning. - Memory provides the facts, context, and history - It should be fast, accurate, and relevant - It should surface what matters, not everything

Never own it. - The system should never claim to "know" something - It should never make a claim without attribution - It should always be able to cite its sources

Vision¶

V3: Reliable Memory¶

Store and retrieve memories reliably
Multiple retrieval strategies (FAISS, BM25, Graph)
Feedback loop for improvement
Local-first, no cloud dependencies

Status: ✅ Complete

V4: Reasoning Infrastructure¶

Goals and planning
Blackboard architecture
Parallel task execution
Computation graphs
Active reasoning over memory

Status: 🚧 In Progress

V5: Cognitive Architecture¶

Hierarchical memory representation
Minimal reconstruction cost
Compression as a side effect
Research phase

Status: 🔬 Research

Eventually: Operating System for Intelligent Agents¶

A complete cognitive architecture
Memory, reasoning, planning, execution
Self-improving over time
Fully local, fully private

Status: 🔮 Vision

Core Principles¶

1. Local First¶

No cloud dependencies
Runs on your hardware
Data stays on your machine
Privacy by default

2. LLM Agnostic¶

Swap models without changing the system
Works with any LLM (Mistral, Llama, GPT, etc.)
The LLM is a plugin, not the core

3. Observable¶

Every decision is traceable
Flight recorder for debugging
Diagnostics and metrics

4. Extensible¶

Pluggable components
Custom signals and strategies
Easy to add new retrieval methods

5. Efficient¶

150ms query latency on 4GB RAM
CPU-only inference
Optimized for local hardware

What We Are Not¶

We Are Not	Because
A vector database	Vector search is one part of memory, not the whole
A chat UI	We provide memory, not a chat interface
A replacement for your brain	Memory is a tool, not a person
A cloud service	Local-first, always
A product	We're building a system, not a product

What We Are¶

A memory system for intelligent agents
A research platform for cognitive architecture
A local-first alternative to cloud memory
An extensible framework for memory experiments

The Big Idea¶

Most AI systems treat memory as an afterthought.

We treat memory as the foundation.

Everything else is built on top.