memory that thinks out loud
Atulya writes its own story
Every post is a retained observation — agents, memory, and the ideas that keep compounding.
Give your OpenAI app a memory in five minutes
Build a ChatGPT-style loop with retain(), recall(), and reflect(). No vector DB, no embedding pipeline, no custom RAG.
What's new in Atulya 0.8.0
New features and improvements in Atulya 0.8.0
Persistent agent memory over MCP
Give any MCP client structured long-term memory: retain, recall, reflect, and mental models. One Docker command for local stack, or Atulya Cloud over HTTPS.
Give any LiteLLM provider persistent memory
atulya-litellm intercepts LiteLLM completions: recall before the call, retain after. Works across OpenAI, Anthropic, Groq, Azure, Bedrock, Vertex, and 100+ providers with one bank.
PydanticAI, observation scopes, and entity labels
Memory tools for PydanticAI agents, tag-scoped consolidation, structured entity labels, timeless retain, and large-bank performance work.
CrewAI agents that remember between runs
CrewAI's built-in memory works inside one kickoff(). When the process exits, context is gone. atulya-crewai implements CrewAI's Storage interface on top of Atulya banks so knowledge compounds across runs.
Integrations hub: Chat SDK, CrewAI, MCP tools, and retrieval upgrades
Model leaderboard, chat middleware, CrewAI storage, configurable MCP tools, batch consolidation, ZeroEntropy reranker, and observation source facts.
Your Vercel Chat SDK bot forgets everything. Atulya fixes that.
TL;DR
File ingestion, batch retain, Go SDK, and DiskANN
Retain PDFs and Office docs directly, cut async retain cost with provider batch APIs, use the Go client, and scale vector search with DiskANN backends.
Production config, observability, and MCP mental models
Per-bank settings, LiteLLM SDK mode, OpenTelemetry, expanded Postgres search, and full mental model lifecycle over MCP.