Local-first cognitive memory

Your AI has amnesia.
We fix that.

A persistent, biologically-inspired memory system for AI agents. Hebbian learning. Energy decay. Semantic retrieval. One SQLite file. No cloud. No vendor lock-in.

Get started — free How it works →
$ curl -fsSL https://memoryco.ai/install.sh | sh
1 file
Your entire brain is a single brain.db
0 ms
Cold start. No containers, no services.
4 states
Active → Dormant → Deep → Archived
∞ recall
Archived memories resurrect on demand
Not another vector database
Memory that thinks like you do
Most "AI memory" is a glorified search index. Memoryco is a cognitive system modeled on how biological memory actually works — with associative recall, natural decay, and learning that strengthens with use.
Typical AI memory

Flat storage

Every memory weighs the same forever. No relevance, no decay, no learning. Just an ever-growing pile of embeddings.

Memoryco

Energy-based decay

Memories have energy that decays over time. Unused memories sink into deep storage. Important memories stay active through natural recall patterns.

Typical AI memory

Isolated retrieval

Search returns results. That's it. No association, no reinforcement, no emergent structure.

Memoryco

Hebbian learning

Memories recalled together form associations automatically. Your knowledge graph builds itself through use — neurons that fire together wire together.

Typical AI memory

Cloud-dependent

Your memories live on someone else's server. Latency, vendor lock-in, privacy concerns, monthly API bills.

Memoryco

Local-first SQLite

One file. Your disk. Zero network calls for memory operations. Back it up, move it, own it. The SQLite of AI memory.

Architecture
Three cognitive layers
Inspired by human cognition — identity is bedrock, references are learned knowledge, engrams are experience. Each layer has different persistence and retrieval characteristics.
Identity
Never decays

Who you are

Persona, traits, values, preferences, relationships, instructions. Always loaded, always available. The bedrock that everything else is built on.

persona traits values preferences relationships instructions
Reference
Immutable source

What you know

Authoritative external knowledge — indexed, searchable, citable. Immutable source documents with a personal annotation layer on top. Full-text FTS5 search with APA 7 citations.

FTS5 search APA 7 cite annotations marketplace
Engram
Decays & learns

What you've experienced

Episodic and semantic memories with energy-based lifecycle. Hebbian association, vector similarity search, and four decay states. Memories never delete — they sink, and can always be resurrected.

Hebbian decay vector search resurrection associations
Built for MCP

Drop it into any agent

Memoryco runs as an MCP server. Add it to Claude Desktop, Cursor, your custom agent — anything that speaks the Model Context Protocol gets persistent memory in one config line.

Your AI stores what it learns, recalls what's relevant, and builds associations through natural conversation. No code changes to your agent. Just memory.

Supports stdio and network transports. SQLite for personal use, Postgres for teams. Same API, same cognitive model.

claude_desktop_config.json
{ "mcpServers": { "memory": { "command": "memoryco", "args": ["--stdio"], "env": { "MEMORY_HOME": "~/.local/share/memory" } } } } // That's it. Your AI now remembers.
Capabilities
Everything you need, nothing you don't
The full cognitive toolkit, from atomic memory storage to graph-based association networks.
engram_*

Engram operations

Create, search, recall, delete, associate. Semantic vector search finds memories by meaning, not just keywords. Batch operations for efficiency.

identity_*

Identity system

Persistent persona, traits, values, preferences, relationships. Always loaded, never decays. The foundation every response builds on.

reference_*

Reference engine

Index authoritative documents with full-text search. APA 7 citations. Immutable sources with personal annotation layers. Pro & Business tiers.

lenses_*

Lenses

Task-specific context guides loaded whole into working memory. Style guides, review checklists, documentation standards. Apply consistently across sessions.

plan_*

Plans

Lightweight multi-step task tracking. Create plans, add steps, mark completion. Transient working documents that keep complex tasks on track.

config_*

Tunable cognition

Adjust decay rate, Hebbian learning rate, recall strength, propagation damping. Every cognitive parameter is a knob you can turn.

A learning system that happens to have an LLM interface, not an LLM that happens to have a knowledge base.

We believe your AI's memory should live on your hardware, decay like biological memory, and learn through the same associative patterns that make human cognition powerful.

No cloud. No API bills. No data harvesting. Just a single SQLite file that gets smarter every time you use it.

1

Local-first, always

Your brain.db lives on your disk. Back it up with cp. Move it with scp. It's a file, not a service.

2

Decay is a feature

Unused memories fade naturally. What remains is what matters. This is how real memory works — and it's how AI memory should too.

3

Use creates structure

Hebbian learning means your knowledge graph emerges from actual usage patterns. No manual tagging, no taxonomy design. Just recall.

4

Nothing is lost

Archived memories aren't deleted. They sink into cold storage and can be resurrected with a single recall. Every experience is recoverable.

Give your AI a brain

Free for personal use. No credit card. No cloud account. Just memory.

See plans Read the architecture →