# Design Reliable Agent Memory Systems

Agents lose useful context across interactions and may retrieve stale or irrelevant memories. This skill provides architectures, storage patterns, retrieval controls, decay methods, and evaluation guidance.

## Install

```bash
npx skillstore add sickn33/agent-memory-systems
```

## Metadata

- Status: approved
- Slug: sickn33-agent-memory-systems
- Skillstore revision: r2
- Version status: missing
- Tree hash: a746e8e4ddbb89f810bc59eeaeb683d53a884aa17936e09df1ec4d28e052ad5a
- Author: sickn33
- GitHub username: sickn33
- License: MIT
- Repository: https://github.com/sickn33/antigravity-awesome-skills/tree/main/skills/agent-memory-systems
- Ref: ebdfe608f5de2b66ff37ab4af12af8ac4f5e8006
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network, env\_access
- Quality score: 76
- Public page: https://skillstore.pages.dev/skills/sickn33-agent-memory-systems
- Manifest: https://skillstore.pages.dev/api/skills/sickn33-agent-memory-systems/manifest

## Capabilities

- Classifies agent information as semantic, episodic, procedural, working, or short-term memory.
- Compares Pinecone, Qdrant, Weaviate, ChromaDB, and pgvector for persistent memory storage.
- Explains fixed-size, semantic, structure-aware, contextual, and code-aware chunking strategies.
- Designs retrieval with metadata filters, hybrid search, reranking, recency scoring, and token budgets.
- Provides patterns for background formation, consolidation, conflict handling, decay, and embedding migration.
- Defines validation checks for isolation, metadata, retrieval limits, and embedding-model consistency.

## Use Cases

- Plan a Persistent Assistant: Define memory types, storage boundaries, retrieval filters, and update rules for an assistant that remembers across sessions.
- Improve Retrieval Quality: Select chunking, metadata, reranking, recency, and evaluation methods for documents and conversation history.
- Review Memory Architecture: Identify stale-memory, context-budget, model-migration, and cross-user leakage risks before production deployment.

## Prompt Templates

### Choose Memory Types

```
Explain semantic, episodic, procedural, working, and short-term memory for [agent]. Recommend what each type should store and explain why.
```

### Design the Architecture

```
Design a memory architecture for [application] with [users] and [data scale]. Compare suitable stores, define namespaces, and include user isolation.
```

### Plan Retrieval and Chunking

```
Create a retrieval and chunking plan for [content types] and [query patterns]. Include filters, chunk sizes, overlap, reranking, token budgets, and evaluation metrics.
```

### Audit a Production Design

```
Audit this memory design: [design]. Identify privacy, cross-user leakage, stale-memory, conflict, context-window, and embedding migration risks. Prioritize fixes and propose validation tests.
```

## Limitations

- Provides architecture guidance and illustrative snippets, not a complete deployable memory service.
- Does not benchmark vector stores or embedding models against your data and workload.
- Requires environment-specific privacy, retention, consent, security, and compliance decisions.
- Does not replace production testing for retrieval quality, latency, cost, or cross-user isolation.

## Best Practices

- Filter every memory query by tenant and user identifiers before applying similarity search.
- Measure retrieval quality with representative queries before selecting chunk sizes, stores, or embedding models.
- Track timestamps, memory type, source, importance, and embedding version in metadata.

## Anti Patterns

- Do not store every interaction forever without consent, retention rules, consolidation, or decay.
- Do not rely on semantic similarity alone when tenant, user, type, or time filters are available.
- Do not mix query and document embeddings from different models or untracked versions.

## Security Audit

- Audited at: 2026-07-23T20:20:32.13\+00:00
- Summary: All 12 static findings are false positives caused by educational examples, ordinary prose, or Markdown formatting. No executable payload, credential exposure, malicious network activity, or prompt injection was found in the skill.

## Stats

- Views: 127
- Downloads: 12
- Favorites: 0
- Popularity score: 0