emk - Episodic Memory Kernel

Part of Agent OS - Kernel-level governance for AI agents

An immutable, append-only ledger for agent experiences. Part of the Agent OS ecosystem.

---

Why emk?

Agents that cannot remember their past repeat the same mistakes. Traditional databases let you update or delete records—this is dangerous for agent memory. What if a bug overwrites critical learning? What if you need to audit agent decisions?

We built emk to subtract the problem of mutable memory. Instead of complex database transactions and state management, we provide a single abstraction: the immutable Episode. Once written, never changed. This constraint eliminates entire classes of bugs and enables reliable agent learning systems.

Scale by Subtraction: By removing the ability to modify history, we remove the need for complex versioning, rollback logic, and concurrency controls. The result is a memory system that scales reliably.

---

Installation

pip install emk

---

Quick Start

from emk import Episode, FileAdapter

store = FileAdapter("agent_memory.jsonl")
episode = Episode(goal="Query user data", action="SELECT * FROM users", 
                  result="200 rows", reflection="Query was fast")
store.store(episode)

That's it. Five lines. Your agent's experience is now permanently recorded.

---

Memory Management Features 🆕

emk now includes powerful memory management features to address the challenge: Infinite memory leads to infinite retrieval costs and confusing context.

Memory Decay & Compression (Sleep Cycle)

Automatically summarize old episodes into semantic rules:

from emk import MemoryCompressor

compressor = MemoryCompressor(store, age_threshold_days=30)
result = compressor.compress_old_episodes()
# Compress 1000 episodes → 20 semantic rules

Benefits: 50-100x storage reduction, faster retrieval, preserved knowledge without overwhelming detail.

Negative Memory (Anti-Patterns)

Explicitly track failures to avoid repeating mistakes:

# Mark failures
failed = episode.mark_as_failure(reason="Connection timeout")
store.store(failed)

# Query both successes AND failures
patterns = store.retrieve_with_anti_patterns()
print(f"What works: {len(patterns['successes'])}")
print(f"What to avoid: {len(patterns['failures'])}")

Benefits: Prune search space instantly, learn from both successes and failures, "DO NOT TOUCH" vectors.

📖 Full documentation: See docs/MEMORY_MANAGEMENT.md and examples/memory_features_demo.py

---

Architecture

emk sits at Layer 1 (Primitives) of the Agent OS stack. It provides the foundational storage abstraction that higher layers depend on:

• Layer 1 (Primitives): emk stores raw episodes. caas (Context-as-a-Service) reads from emk to build working memory. cmvk (Cryptographic Message Verification Kernel) may sign episodes for audit trails.
• Layer 2 (Infrastructure): amb (Agent Message Bus) transports messages. iatp (Inter-Agent Trust Protocol) verifies trust. These layers may log their own episodes into emk.
• Layer 3 (Framework): agent-control-plane orchestrates agents. scak (Self-Correction and Alignment Kernel) uses emk to learn from mistakes.

emk has zero dependencies on other Agent OS components. It is the foundation, not the application.

---

The Ecosystem Map

emk is one component of the modular Agent OS. Here's how the pieces fit together:

Layer	Component	Purpose	Repository
Primitives	caas	Context-as-a-Service (Working Memory)	caas
Primitives	cmvk	Cryptographic Verification Kernel	cmvk
Primitives	emk	Episodic Memory Kernel	emk
Infrastructure	iatp	Inter-Agent Trust Protocol	iatp
Infrastructure	amb	Agent Message Bus	amb
Infrastructure	atr	Agent Tool Registry	atr
Framework	agent-control-plane	Agent Orchestration and Control	agent-control-plane
Framework	scak	Self-Correction and Alignment Kernel	scak

Each component is independently useful. Together, they form a complete Agent Operating System.

---

Citation

If you use emk in your research or production systems, please cite:

@software{emk2024,
  author       = {Siddique, Imran},
  title        = {emk: Episodic Memory Kernel for Agent Systems},
  year         = {2024},
  publisher    = {GitHub},
  url          = {https://github.com/imran-siddique/emk},
  note         = {Part of the Agent OS ecosystem}
}

---

License

MIT License - see LICENSE for details.