Please read this Version file
███╗ ███╗███████╗███╗ ███╗███╗ ███╗ ██████╗ ██╗ ████████╗
████╗ ████║██╔════╝████╗ ████║████╗ ████║██╔═══██╗██║ ╚══██╔══╝
██╔████╔██║█████╗ ██╔████╔██║██╔████╔██║██║ ██║██║ ██║
██║╚██╔╝██║██╔══╝ ██║╚██╔╝██║██║╚██╔╝██║██║ ██║██║ ██║
██║ ╚═╝ ██║███████╗██║ ╚═╝ ██║██║ ╚═╝ ██║╚██████╔╝███████╗██║
╚═╝ ╚═╝╚══════╝╚═╝ ╚═╝╚═╝ ╚═╝ ╚═════╝ ╚══════╝╚═╝
MemMolt
Structured, searchable, long-term memory for your AI agent. Plugs into Claude Code (and any MCP-compatible client) over the Model Context Protocol.
What is MemMolt?
MemMolt is a memory system for AI agents. Think of it as a brain extension for Claude Code that remembers things across conversations, keeps them tidy, and can recall them by meaning — not just keywords.
It lives on your machine as a tiny local server. Your agent connects to it, reads from it, writes to it, and picks up where it left off the next time you open a conversation.
Memory is organized into three simple layers:
BUCKET → top-level category ("Personal Finances")
│
└── THREAD → sub-topic ("Quarterly Taxes")
│
└── MEMO → the actual note ("Q3 estimated payment notes")
That's it. Buckets hold threads. Threads hold memos. Memos are markdown documents. No nested folders, no wiki rabbit holes, no broken links.
What problem does it solve?
If you've used an AI agent for a while, you've probably hit one of these:
- "It forgot what we talked about yesterday."
- "I keep re-explaining the same context every session."
- "My notes folder has grown into a jungle I can't navigate."
- "The agent wastes half its context window reading old notes."
MemMolt fixes all four. The agent only pulls in the memos it actually needs, organized so it knows where to look, indexed so it can find them by meaning and not just word match.
Why not just use Claude Code + Obsidian (or any notes folder)?
A plain notes folder works — until it doesn't. Here's what you trade away:
| Problem | Plain notes folder | MemMolt |
|---|---|---|
| Finding things | Text search only. Miss the exact word, miss the note. | Hybrid search: keyword (FTS5) + meaning (vector embeddings) combined via RRF so you get both. |
| Organization | Folders can be nested arbitrarily, links break, stuff drifts. | Enforced 3-level hierarchy. The agent can't make a mess because the structure doesn't allow one. |
| Context window | Agent reads multiple full files just to check if they're relevant. | Agent searches summaries first, fetches only the memos it actually needs. Massively fewer tokens. |
| Spiraling out | Long notes get longer, topics sprawl across files. | Summaries force compression. Each memo has a forced title + summary that must describe what it contains. |
| Consistency | The agent might remember to update notes, or it might not. | The tool prompts the agent to update summaries after changes. Built-in nudges keep memory fresh. |
| Search across topics | Manual. You grep, you browse, you hope. | One call returns results from the whole system, ranked by relevance. |
Short version: A notes folder is a filesystem. MemMolt is a memory system.
How it works (the quick version)
- Everything lives in one SQLite file on your disk (default:
~/.memmolt/memmolt.sqlite— see Configuration for the full resolution order). - Summaries are turned into vectors by a local embedding model (all-MiniLM-L6-v2, runs in-process, no cloud calls).
- Search combines keyword matching (FTS5) and semantic matching (sqlite-vec), then merges them with Reciprocal Rank Fusion (RRF).
- The agent talks to MemMolt over MCP (Model Context Protocol) — the same way Claude Code talks to any other tool.
No separate database server. N