Notre avis

Atlas ingère des fichiers de documentation et de code dans une mémoire sémantique persistante en utilisant Qdrant et les embeddings Voyage pour permettre une recherche et un rappel de contexte entre sessions.

Points forts

Mémoire persistante entre sessions, éliminant le problème de dépassement de contexte.
Recherche sémantique performante avec dual-indexing (clés sémantiques et temporelles).
Ingestion récursive de dossiers avec support de multiples types de fichiers.
Faible latence (10-50 ms) et haute précision de rappel (Recall@10 > 0.98).

Limites

Nécessite Qdrant en cours d'exécution localement et une clé API Voyage.
Consommation mémoire importante (1,4 Go RAM) pour de gros volumes de vecteurs.
Ne prend en charge que les types de fichiers listés (pas d'images, PDF, etc.).

Quand l'utiliser

Utilisez Atlas lorsque vous devez conserver un contexte volumineux entre plusieurs sessions ou rendre une base de documentation interrogeable sémantiquement.

Quand l'éviter

Ne l'utilisez pas si votre projet est très petit ou si vous n'avez pas besoin de mémoire persistante entre sessions.

Exemples

Remember architecture decision

Remember the API architecture we discussed and store it in Atlas so I can retrieve it later.

Ingest project documentation

Ingest all markdown files from the docs/ folder recursively into Atlas.

Search previous work

Search Atlas for all mentions of authentication patterns in the codebase.

name: atlas description: Ingest project documentation and code into persistent semantic memory (Qdrant + Voyage embeddings). Use when user wants to remember context across sessions, ingest docs, or search previous work. Requires Qdrant running locally and VOYAGE_API_KEY set. allowed-tools: Bash(bun:*)

Atlas - Persistent Semantic Memory

Atlas provides automatic context ingestion and retrieval using Voyage embeddings + Qdrant vector database. Solves the context overflow problem by storing knowledge persistently across sessions.

Quick Start

Prerequisites

Qdrant running locally:

docker run -d -p 6333:6333 qdrant/qdrant

VOYAGE_API_KEY set (get from https://voyageai.com):

export VOYAGE_API_KEY="your-key-here"

Verify setup:

curl http://localhost:6333/health

Ingesting Context

Store files in Atlas memory for persistent retrieval:

Ingest Single File

cd ~/production/atlas
bun run --filter @inherent.design/atlas atlas ingest /path/to/file.md

Ingest Directory (Recursive)

cd ~/production/atlas
bun run --filter @inherent.design/atlas atlas ingest /path/to/docs/ --recursive

Ingest Multiple Paths

cd ~/production/atlas
bun run --filter @inherent.design/atlas atlas ingest README.md src/index.ts docs/ -r

What gets ingested:

Supported: .md, .ts, .tsx, .js, .jsx, .json, .yaml, .qntm, .rs, .go, .py, .sh, .css, .html
Ignored: node_modules, .git, dist, build, coverage, .atlas

Processing:

Chunks text (768 tokens, 13% overlap) for semantic coherence
Embeds with Voyage-3-large (1024-dim)
Stores in Qdrant with dual-indexing (semantic QNTM keys + temporal timestamps)
Preserves original text for future consolidation

Searching Context

Retrieve relevant context semantically:

Basic Search

cd ~/production/atlas
bun run --filter @inherent.design/atlas atlas search "typescript patterns"

Limited Results

cd ~/production/atlas
bun run --filter @inherent.design/atlas atlas search "memory consolidation" --limit 10

Temporal Filtering (Since Date)

cd ~/production/atlas
bun run --filter @inherent.design/atlas atlas search "sleep patterns" --since "2025-12-25"

Chronological Timeline

cd ~/production/atlas
bun run --filter @inherent.design/atlas atlas timeline --since "2025-12-01"

When to Use This Skill

Use Atlas when:

User asks to "remember this across sessions"
Project context is too large for single session
User wants to search previous work/decisions
Documentation needs to be queryable
Building on previous research or code

Examples:

"Remember the API architecture we discussed"
"What did we decide about the database schema?"
"Find all mentions of authentication patterns"
"Ingest all the .atlas research files"

Architecture

Built on .atlas research (Steps 1-4 + Sleep Patterns):

Stack:

Voyage-3-large embeddings (1024-dim, 9.74% better than OpenAI)
Qdrant HNSW index (M=16, int8 quantization, 4x compression)
RecursiveCharacterTextSplitter (semantic boundaries)
Dual-indexing (QNTM semantic keys + RFC 3339 timestamps)

Production Config (from Step 3 research):

Recall@10: >0.98
Latency: 10-50ms (p95)
Memory: 1.4GB RAM + 5GB disk per 1M vectors

Technical Details

For implementation details, see:

docs/architecture.md - Complete technical architecture

Packages:

@inherent.design/atlas-core - Core library (embeddings, storage, search)
@inherent.design/atlas - Command-line interface
@inherent.design/atlas-mcp - MCP server for Claude Code integration

Atlas - Mémoire sémantique persistante

Recommandé pour