LLMFS — Filesystem Memory for LLMs and AI Agents¶

LLMFS gives LLMs and AI agents persistent, searchable, structured memory — organized like a filesystem. Instead of losing context when a conversation grows past the token limit, agents offload memories to LLMFS and retrieve exactly what they need, when they need it.

The result: zero information loss and an effectively unlimited context window — even over thousands of turns.

The Problem¶

Every LLM agent eventually hits the same wall: the context window fills up.

The standard solution — lossy summarization — destroys information. When an agent summarizes 80k tokens into 5k, 94% of the detail is gone forever. Ask it about a specific line of code from 30 turns ago, and it can only apologize.

The Solution¶

LLMFS takes a different approach, borrowed directly from operating systems:

OS Concept     →   LLM Concept
──────────────────────────────────────────────────────────
RAM            →   Context Window (e.g. 128k tokens)
Disk / Swap    →   LLMFS  (500k+ tokens, full fidelity)
Page eviction  →   Offload old turns to LLMFS
Page fault     →   LLM calls memory_search / memory_read
Virtual addr   →   Memory path  (/session/turns/42)
MMU            →   ContextManager

Memories are stored at filesystem-style paths (/projects/auth/bug, /events/2026-03-15_fix) and searched semantically. They persist across sessions, support TTLs, carry metadata and tags, and can be linked in a knowledge graph.

Quick Start¶

pip install llmfs
llmfs init
llmfs write /knowledge/hello "LLMFS stores memories at filesystem paths"
llmfs search "how does memory storage work"

from llmfs import MemoryFS

mem = MemoryFS()
mem.write("/projects/auth/bug", "JWT expiry misconfigured at auth.py:45", tags=["jwt", "bug"])
results = mem.search("authentication error", k=3)
print(results[0].path, results[0].score)

Get Started Python API GitHub

Key Features¶

Performance¶