prompt

Free

Design memory systems for long-running AI agents

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About prompt

Agent Memory Architect is a comprehensive prompt for designing memory systems in long-running AI agents. It provides detailed guidance on short-term and long-term memory design, including compression strategies, summarization triggers, and pruning rules for STM, as well as extraction, storage, retrieval, and update mechanisms for LTM. It covers four memory types (episodic, semantic, procedural, metacognitive), integration with reasoning through thought retrieval, the full memory lifecycle (write path, read path, maintenance), and observability features such as drift detection and hit/miss rate tracking. The prompt is research-backed, citing multiple academic papers, and emphasizes structured memory over free text.

Key Features

Design short-term memory with context-window budget, compression, and pruning

Design long-term memory with extraction, storage (vector DB, knowledge graph), and retrieval

Choose memory types: episodic, semantic, procedural, metacognitive

Integrate memory with reasoning using thought retrieval and uncertainty handling

Define memory lifecycle: write path, read path, maintenance (deduplication, aging)

Ensure observability with retrieval tracking, hit/miss rates, and drift detection

Pros & Cons

Pros

Comprehensive coverage of memory system design principles
Research-backed with citations to multiple academic papers
Emphasizes explicit, inspectable, and bounded memory operations
Includes actionable design principles and lifecycle management
Free and open-source prompt available on GitHub

Cons

Requires manual implementation of the described memory system; no code provided
Only provides design guidance, not a turnkey solution
May require significant LLM engineering effort to realize the architecture

Best For

Building long-running conversational agents that learn from experienceDeveloping AI agents that avoid repetitive mistakes and retrieve relevant contextImplementing memory-augmented reasoning in large language modelsDesigning modular memory systems for research and experimentationCreating agents with bounded memory to manage token budgets

FAQ

What is the purpose of the Agent Memory Architect prompt?

It is designed to help build memory systems that let long-running agents learn from experience, avoid repetitive mistakes, and retrieve the right context at the right time without token bloat or stale recollections.

What memory types does this prompt cover?

It covers episodic (specific events), semantic (general facts), procedural (skills and patterns), and metacognitive (confidence and failure modes) memory types.

Does this prompt include short-term and long-term memory?

Yes, it includes design for short-term memory (context window budget, compression, pruning) and long-term memory (extraction, storage, retrieval, update/delete).

What is the emphasis on observability?

The prompt stresses tracking what was retrieved, why, and how it influenced the response, as well as memory hit/miss rates and drift detection to detect outdated or harmful memories.