What is memledger?

MemLedger is a memory system for AI agents that keeps a full, traceable history of every fact the agent remembers, so you can always find out where a memory came from and why it was kept. Most memory tools let an AI chatbot remember things across conversations, but once a fact is stored you usually cannot tell which sentence it was pulled from, which model extracted it, or why an old preference suddenly disappeared. MemLedger solves that by logging every step: a single command, memledger why, shows the exact sentence a memory was born from, the model that extracted it, when it became a permanent fact, and who approved that promotion. The system stores memory in three layers that mirror how people remember things. Instinct holds core facts that get included in every conversation, either set by you directly or promoted automatically once a fact proves reliable across several sessions. Episodic holds longer term facts extracted by an AI model at checkpoints, each with a confidence score and an expiration time. Working memory holds the current conversation's turn by turn history. All three layers are built on top of a single append only log of events stored in SQLite, so the log itself, not any summary view, is treated as the true record. To keep costs down, MemLedger uses a free, rule based filter to decide which parts of a conversation are worth sending to an AI model for fact extraction at all, skipping small talk that carries no real information. New facts are also quarantined and need to be confirmed across multiple sessions, or approved by a person, before becoming permanent, and if a bad fact does slip through, its full history lets you trace it and delete it along with anything that was derived from it. Because the raw conversation history is always kept, you can regenerate the entire memory later with a better AI model instead of being stuck with whatever an earlier, weaker model extracted. It runs as a single SQLite file on your own machine, works with a local model through Ollama or with any cloud AI service, requires no server, and is released under the MIT license.