developersglobal/ai-agent-skills/rag-and-memory

Overview

RAG and memory systems are how AI agents work with knowledge that exceeds their context window. Done well: agents give accurate, grounded answers. Done poorly: context overflow, hallucination from stale retrieval, and performance degradation.

This skill covers the design principles and failure modes of RAG and memory architectures for production AI systems.

When to Use

• Building any AI system that needs to access external knowledge
• When agent context windows are being exceeded
• When agents need to remember information across sessions
• When building Q&A, document analysis, or knowledge base systems

Process

Step 1: Choose the Right Memory Architecture

1. Identify what the agent needs to remember:

• Ephemeral: Within a single session (use in-context memory)
• Session-persistent: Across a user's sessions (use external key-value store)
• Knowledge base: Organizational or domain knowledge (use vector DB + RAG)
• Procedural: How to do tasks (encode in SKILL.md / system prompt)

1. Match the memory type to the store:

Verify: Each type of information the agent needs has a defined storage mechanism.

Step 2: Design the RAG Pipeline

1. Chunking strategy: Break documents into chunks at semantic boundaries (paragraphs, sections) — not arbitrary character counts.
2. Embedding model: Match the embedding model to your query type. Use the same model for indexing and retrieval.
3. Retrieval: Retrieve top-K most semantically similar chunks. K = 3–7 is usually optimal.
4. Re-ranking: After retrieval, re-rank by relevance using a cross-encoder. Top K becomes top 3–5 for the prompt.
5. Context injection: Inject retrieved chunks into the prompt with clear source citations.

Verify: Retrieved chunks are genuinely relevant to the query before injecting into context.

Step 3: Prevent Context Bloat

1. Summarize, don't accumulate: For long sessions, summarize previous turns rather than appending them indefinitely.
2. Retrieve, don't pre-load: Only load context relevant to the current query. Don't pre-load everything.
3. Set context budgets: Define maximum token allocations for: system prompt, retrieved context, conversation history, user message.
4. Compress before injecting: Summarize long retrieved documents to extract the relevant portion only.

Verify: Total prompt length is within model limits with buffer. Retrieved context is relevant to current query.

Step 4: Handle Retrieval Failures Gracefully

1. If retrieval returns no relevant results: say so — do not hallucinate an answer.
2. If retrieved documents are outdated: surface the document date to the user.
3. If confidence is low: present the retrieved source and let the user evaluate.
4. Design for "no relevant information found" as a first-class outcome.

Verify: System has defined behavior for failed/empty retrieval.

Step 5: Measure and Optimize

1. Track retrieval quality:

• Precision: Are retrieved chunks relevant to the query?
• Recall: Are relevant chunks being retrieved at all?

1. Track answer quality: Use RAGAS or similar evaluation framework.
2. Monitor: context length per query, retrieval latency, hallucination rate.

Verify: Baseline metrics established. Retrieval precision > 80%.

Common Rationalizations (and Rebuttals)

Red Flags

• Entire document corpus pre-loaded into every prompt
• Retrieval returning chunks from unrelated documents
• No defined behavior for empty retrieval results
• Context window regularly at 90%+ capacity
• Agent answering from "training knowledge" instead of retrieved documents
• No source citations for retrieved information

Verification

• [ ] Memory architecture matches the type of information needed
• [ ] RAG pipeline: chunk → embed → retrieve → re-rank → inject
• [ ] Context budgets defined for all prompt sections
• [ ] Empty retrieval has a defined graceful fallback
• [ ] Retrieval precision measured and > 80%
• [ ] Source citations included in AI responses

References

• hallucination-prevention skill
• multi-agent-orchestration skill
• ai-output-validation skill