mnemonic-ai 0.1.0

Framework-agnostic memory layer for AI agents with importance-aware, graph-structured context assembly

Mnemonic

Give your AI agents a memory that works like a brain.

Today's AI agents are stateless. Every conversation starts from zero. Mnemonic changes that. It's a memory layer that models how humans actually remember — important things stick, old things fade, related ideas connect, and contradictions get flagged.

Tell your agent you prefer Kubernetes on Monday, and it still knows on Friday. Correct a mistake, and the old fact fades while the new one takes priority. No more repeating yourself.

How It Works

Mnemonic organizes memories into five tiers, just like the human brain:

Tier	What it holds	How long it lasts
Identity	Name, role, core preferences	Years
Procedural	Workflows, step-by-step processes	~1 year
Structural	Architecture decisions, tools, constraints	~6 months
Episodic	Specific conversations and events	~30 days
Transient	Small talk, temporary context	~7 days

Memories don't just sit in a list. They form a graph — related ideas link together, contradictions are detected, and outdated facts get superseded (not deleted). When your agent needs context, Mnemonic picks the most relevant memories that fit within the token budget.

Get Started in 3 Lines

pip install mnemonic-ai

from mnemonic import Mnemonic

mem = Mnemonic()
mem.add("User prefers dark mode and Kubernetes over ECS")
context = mem.recall("deployment preferences", max_tokens=3000)
# Returns the most relevant memories, ranked by importance

No API keys. No database setup. Works immediately with SQLite.

Why Mnemonic

• Memories decay naturally — like human memory, unused information fades. Actively accessed memories stay strong.
• Smart prioritization — urgency, errors, corrections, and decisions automatically get importance boosts.
• Contradictions are handled — "moved to SF" supersedes "lived in NYC" instead of creating confusion.
• Stays within your token budget — fills the context window with the highest-signal memories, not everything.
• Gets smarter over time — episodic memories consolidate into durable knowledge, just like sleep does for the brain.

Scale When You're Ready

What you need	How to get it
Production database	pip install mnemonic-ai[postgres] — PostgreSQL with vector search
Semantic search	pip install mnemonic-ai[openai-embeddings] or mnemonic-ai[local-embeddings]
AI tool integration	pip install mnemonic-ai[mcp] then mnemonic-mcp — works with Claude, VS Code Copilot
TypeScript	npm install @mnemonic-ai/core — full parity with Python
Multiple agents	Built-in agent isolation with explicit sharing

Use With Any Framework

# LangChain
from mnemonic.adapters.langchain import LangChainMemory

# CrewAI
from mnemonic.adapters.crewai import CrewAIMemory

# AutoGen
from mnemonic.adapters.autogen import AutoGenMemory

# Or just inject into any prompt
from mnemonic.adapters.raw import recall_as_prompt
system = f"You are helpful.\n\n{recall_as_prompt(mem, user_query)}"

Multi-Agent Memory

Multiple agents can share a memory store while keeping their own memories private:

agent_a = Mnemonic(store=store, agent_id="researcher")
agent_b = Mnemonic(store=store, agent_id="coder")

agent_a.add("The API uses OAuth 2.0", shared=True)  # both agents see this
agent_b.add("Refactored auth module yesterday")       # only coder sees this

Extend It

Mnemonic is designed to be extended. Register custom store backends, add consolidation hooks, or replace any heuristic with your own logic:

from mnemonic import register_store, MnemonicConfig

register_store("mydb", my_factory)        # custom store backend

cfg = MnemonicConfig()
cfg.scorer.weight_recency = 0.50          # tune scoring weights
cfg.classifier.custom_classifier = my_fn  # replace tier classification

Mnemonic Platform

Need Neo4j graph storage, LLM-powered consolidation, a REST API with authentication, or an admin dashboard? See Mnemonic Platform.

License

Apache 2.0