cogmemory 0.1.0a1

Cognitive memory architecture for LLM agents with principled forgetting

Mnemo

Cognitive memory architecture for LLM agents with principled forgetting.

Mnemo provides a biologically-inspired memory system for AI agents, implementing tiered storage (sensory → working → long-term) with automatic decay and consolidation. Built on Ebbinghaus forgetting curves and modern memory consolidation theory, it lets agents maintain relevant context while gracefully forgetting stale information — just like humans do.

Installation

pip install mnemo

For real embedding support (sentence-transformers):

pip install mnemo[embeddings]

Quick Start

import mnemo

# Create a three-tier memory system
system = mnemo.create_memory_system()

# Store a memory
mem = mnemo.MemoryUnit(content="User prefers concise answers", importance=0.8)
system["working"].append(mem)

# Simulate retrieval (strengthens the memory)
mem.access()

# Check if it should be promoted to long-term storage
model = system["model"]
if model.should_consolidate(mem):
    system["long_term"].append(mem)

How It Works

Mnemo models agent memory as three tiers:

Tier	Purpose	Capacity	Decay
Sensory	Raw input buffer	High	Very fast
Working	Active context	Limited	Moderate
Long-term	Consolidated knowledge	Large	Slow

Memories decay exponentially based on time since last access, modulated by importance and retrieval frequency. Frequently accessed working memories consolidate into long-term storage.

Paper

Principled Forgetting in LLM Agent Memory Systems [Link to paper forthcoming]

License

Apache 2.0 — see LICENSE for details.