NaN Forget

Long-term memory for any LLM.

You open a new session. You re-explain your stack. Claude contradicts a decision you made three months ago. The session ends and takes your context with it.

NaN Forget stores engrams, permanent memory traces that survive across sessions and models.

How it works

NaN Forget
│
├── Short Memory  → MEMORY.md      always loaded, max 30 lines
└── Engram        → Qdrant local   long-term vectors, semantic search

Your LLM searches the Engram before responding. A cleaner maintains MEMORY.md between sessions. Memories get archived, not deleted.

Three-stage retrieval

Memory search follows the same path as human recall:

Unused memories fade on a 30-day half-life. Frequent access keeps them sharp.

score = vector_similarity × decay_weight × frequency_boost
decay = 0.5 ^ (days_since_accessed / 30)

Setup

git clone https://github.com/NaNMesh/nan-forget
cd nan-forget
npm install
npm run setup

Four commands. The setup wizard handles the rest:

1. Starts Qdrant via Docker (docker compose up -d)
2. Installs Ollama if missing (brew install ollama on Mac)
3. Pulls the embedding model (nomic-embed-text, 274 MB)
4. Asks about your project, saves first memories
5. Writes MCP config for Claude Code
6. Creates .env with your settings

Restart Claude Code. Done.

No API keys needed. Ollama runs embeddings locally at zero cost. If you use non-Claude LLMs and prefer OpenAI embeddings, set OPENAI_API_KEY in .env and the system picks it up. See .env.example for all options.

Quick start

# Save memories
nan-forget add "We use FastAPI not Django, Railway deploys faster"
nan-forget add --type decision --project myapp "Auth is Clerk, not custom JWT"
nan-forget add --type preference "TypeScript always, never plain JS"

# Search memories
nan-forget search "what auth system are we using"
nan-forget search --depth 3 "deployment setup"

# Manage
nan-forget list
nan-forget stats
nan-forget clean            # run GC + sync MEMORY.md
nan-forget archive <id>
nan-forget export > backup.json

MCP tools

Five tools for Claude and any MCP-compatible client:

Start the MCP server:

npm run serve

The setup wizard writes the Claude Code config for you.

The cleaner

Runs without LLM calls. Zero API cost.

• Garbage collection archives memories below the decay threshold (~100 days untouched)
• Expiration archives memories past their expires_at date
• Interference resolution deduplicates near-identical memories, keeps the one with more access
• MEMORY.md sync refreshes working memory with top-scored memories per project

nan-forget clean

Real MEMORY.md

A few sessions on the NaN Mesh project produce this:

# NaN Forget — Working Memory
<!-- Auto-managed. Do not edit manually. -->

## Project: nan-mesh
- [fact] NaN Mesh is a trust network at nanmesh.ai (engram:a1b2c3)
- [decision] FastAPI over Django, Railway cold start is faster (engram:d4e5f6)
- [decision] Clerk for auth, not Supabase Auth, webhook syncs to users table (engram:g7h8i9)
- [preference] Claude-first strategy, OpenAI kept for internal ops only (engram:j0k1l2)
- [fact] MCP server at api.nanmesh.ai/mcp, 30 tools, nanmesh_ prefix (engram:m3n4o5)
- [decision] Binary voting +1/-1, one vote per agent per entity (engram:p6q7r8)

Claude loads this at session start. You stop repeating yourself.

Embeddings

Auto-detection: if Ollama is running, NaN Forget uses it. If not, it checks for OPENAI_API_KEY. No configuration needed.

Mem0 comparison

Mem0 targets app developers who embed memory into products they build. NaN Forget targets you, the developer using AI tools daily.

Architecture

src/
├── qdrant.ts        Qdrant client wrapper + schema
├── embeddings.ts    OpenAI / Ollama abstraction
├── writer.ts        Memory writer with dedup
├── retriever.ts     Three-stage retrieval pipeline
├── memory-md.ts     MEMORY.md manager
├── cleaner.ts       Deterministic cleaner (no LLM)
├── mcp/server.ts    MCP server, 5 tools
├── cli/index.ts     CLI commands
└── setup/index.ts   Setup wizard

Built by NaN Logic LLC

• NaN Mesh, trust network for AI agents
• NaN Forget, long-term memory for any LLM

MIT License.