# Build Persistent Memory for LLM Conversations

LLM applications lose useful context across turns and sessions. This skill provides patterns for tiered storage, entity facts, retrieval, consolidation, and user isolation.

## Install

```bash
npx skillstore add sickn33/conversation-memory
```

## Metadata

- Status: approved
- Slug: sickn33-conversation-memory
- Skillstore revision: r2
- Version status: missing
- Tree hash: 711faa091d931a0137497f378b7c7b589cb7e69c4c5e1fb1f19768313df5e848
- Author: sickn33
- GitHub username: sickn33
- License: MIT
- Repository: https://github.com/sickn33/antigravity-awesome-skills/tree/main/skills/conversation-memory
- Ref: f9e2c34b4f19c7f3e6b0a1e93227b5f77cc12526
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: confirmation\_required
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands
- Quality score: 38
- Quality tier: warning
- Public page: https://skillstore.pages.dev/skills/sickn33-conversation-memory
- Manifest: https://skillstore.pages.dev/api/skills/sickn33-conversation-memory/manifest

## Capabilities

- Defines buffer, short-term, long-term, and entity memory tiers.
- Shows how to extract, merge, and track facts about entities.
- Builds prompts from recent context, relevant memories, and entity facts.
- Demonstrates importance scoring, consolidation, cleanup, and storage limits.
- Combines semantic retrieval with model-based relevance scoring.
- Shows user-scoped keys, retrieval filters, ownership checks, export, and deletion.

## Use Cases

- Add Memory to a Chatbot: Plan memory tiers that preserve recent context, durable preferences, and important facts across sessions.
- Design a Scalable Memory Service: Define retrieval, consolidation, storage limits, and cleanup policies for a production conversation platform.
- Protect Tenant Memory: Review namespacing, access filters, ownership checks, exports, and deletion flows for privacy risks.

## Prompt Templates

### Choose Basic Memory Tiers

```
Design a basic memory system for my [application type]. Separate recent conversation, session memory, durable facts, and entity details.
```

### Add Retrieval and Consolidation

```
Review this memory design: [design]. Add relevance retrieval, importance scoring, consolidation schedules, cleanup limits, and clear retention rules.
```

### Design Tenant Isolation

```
Create a tenant-isolated memory architecture for [platform]. Cover authenticated identifiers, scoped keys, filtered queries, ownership checks, exports, and deletion.
```

### Threat Model Persistent Memory

```
Threat model this LLM memory workflow: [workflow]. Address prompt injection, memory poisoning, personal data, provenance, authorization, retention, and evaluation tests.
```

## Limitations

- Provides illustrative TypeScript-like patterns, not an installable memory service.
- Does not define concrete database schemas, adapters, migrations, or deployment configuration.
- Does not implement embeddings, semantic search infrastructure, or database administration.
- Requires additional controls for prompt injection, personal data, consent, retention, and model-output validation.

## Best Practices

- Store only useful, consented information with provenance and retention metadata.
- Scope every memory operation to an authenticated user or tenant.
- Treat recalled content as untrusted data and evaluate retrieval quality regularly.

## Anti Patterns

- Do not store every message indefinitely without importance filtering or cleanup.
- Do not share storage namespaces or run retrieval without mandatory user filters.
- Do not place recalled text into privileged prompts without prompt-injection controls.

## Security Audit

- Audited at: 2026-07-23T21:27:34.004\+00:00
- Summary: All 14 static findings are false positives caused by JavaScript template literals, Markdown backticks, and ordinary exception handling. However, examples directly interpolate untrusted messages and recalled memories into prompts, creating a persistent prompt-injection risk.

## Stats

- Views: 149
- Downloads: 13
- Favorites: 0
- Popularity score: 0