0.9.4

What's Changed

New Features

• Make OpenAI and Anthropic API base URLs configurable by @abrookins in #43

Fixes

• Fix hash-based deduplication FT.AGGREGATE query execution by @abrookins in #41

Full Changelog: server/v0.9.3...server/v0.9.4

0.9.3

What's Changed

• Add remaining context percentage until auto-summarization to working memory endpoints by @abrookins in #38
• Fix authentication event loop corruption by converting get_current_user to async by @abrookins in #40

Full Changelog: server/v0.9.2...server/v0.9.3

0.9.2

What's Changed

• Token-based authentication by @abrookins in #35

Full Changelog: server/v0.9.1...server/v0.9.2

0.9.1

What's Changed

• Fix docker-compose file by @abrookins in #34

Full Changelog: server/v0.9.0...server/v0.9.1

0.9.0

Memory Evolution

• Working Memory (formerly Short-term Memory):

  • Renamed from "short-term memory" to "working memory" to better reflect its purpose
  • Enhanced with automatic promotion system that moves structured memories to long-term storage in background
  • Added support for arbitrary JSON data storage alongside memory structures
  • Improved automatic conversation summarization in working memory, based on token limits

• Long-term Memory Promotion:

  • Implemented seamless flow from working memory to long-term memory via background task processing
  • Agent only has to think about working memory, long-term memory is managed automatically (but can be managed manually, too)
  • Use any LangChain VectorStore subclass for long-term storage, defaults to RedisVectorStore
  • Structured memories are automatically promoted with vector embeddings and metadata indexing
  • Deduplication and compaction systems for long-term memory management
  • Background task worker system using for reliable, scalable memory processing

Client SDK and Tooling

• Working and long-term memory available as tools for LLM integration (LLM can choose to persist a long-term memory or search for long-term memories, etc.)
• Higher-level tools support sending in a user's input and getting back a context-enriched prompt, via /v1/memory/prompt endpoint
• Support for namespace isolation, user separation, and session management

Search and Retrieval

• Vector-based similarity search using OpenAI embeddings
• Rich filtering system by session, namespace, topics, entities, timestamps
• Hybrid search combining semantic similarity with metadata filtering
• RedisVL integration for high-performance vector operations with Redis

Enhanced Memory Classification:

• Semantic memories for facts and preferences
• Episodic memories for time-bound events with event dates (requires a timeframe)
• Message memories for long-term conversation records (optional)
• Automatic topic modeling and entity recognition either using BERTopic or a configured LLM
• Rich metadata extraction and indexing

Authentication and Security

• OAuth2/JWT Bearer token authentication with JWKS validation
• Multi-provider support (Auth0, AWS Cognito, Okta, Azure AD)
• Role-based access control using JWT claims
• Development mode with configurable auth bypass

Operational Features

• Comprehensive CLI Interface:
  • Commands for server management (api, mcp, task-worker)
  • Database operations (rebuild-index)
  • Background task scheduling and management
  • Health monitoring and diagnostics