BioMem:A local-first memory engine for LLM conversations

Most of what AI products call memory is storage. They keep the text and paste back something similar. BioMem is my memory system for AI, built to behave like memory. It has short-term and long-term layers and sleep-style consolidation.

Visit biomem.bleeding.dev Source on GitHub

The open-source engine is alpha 0.0.2 under the MIT license.

AI Does Not Remember. It Only Stores. On stage at Černá AI Festival 2026, Ostrava, 15 Sept 2026

Storage or memory? Try both.

Send the same messages to both sides and see what each one keeps.

Press Play to watch an example, or pick a message.

Each message adds fatigue. When fatigue reaches the limit, sleep runs on its own.

Last step

You said: Releases go out on Friday. Storage appended it and keeps every entry. Its closest earlier match by text: Friday lunch is pizza. Memory put it in short-term. Unused memories faded out.

Storage

Storage keeps every message and returns the one with the most similar text.

Closest match by text

Friday lunch is pizza

Releases go out on Friday
Store dates in UTC
Installs go through pnpm, never npm
Never commit .env files
Friday lunch is pizza
Use pnpm for installs

Other earlier entries in storage: 3

Memory

Fatigue: 2 of 4

Repeated messages merge and grow stronger. Unused ones fade. Sleep moves the strongest short-term memories into long-term.

Short-term, fades fast

Store dates in UTC, Strength: medium
Releases go out on Friday, Strength: medium

Long-term, fades slowly

Installs go through pnpm, never npm, Strength: medium
Never commit .env files, Strength: low

This demo simulates the ideas in the BioMem README. It does not run the BioMem engine.

Where BioMem runs

On your machine. The open-source engine keeps memory databases on your machine and does not sync them. It is pure Python 3.10+ for macOS, Linux and Windows.
MCP and browser extensions. A local stdio MCP adapter works with MCP hosts. Browser extensions are for Chromium and Firefox.
Beta for coding agents. biomem.bleeding.dev has a beta of long-term memory for coding agents such as Claude Code, Codex and OpenCode.

Codebase Radar, an AI engineering platform for production-ready software, is in development.