Persistent AI Memory System v2.0.0

🌟 Community Call to Action: Have you made improvements or additions to this system? Submit a pull request! Every contributor will be properly credited in the final product.

GITHUB LINK - https://github.com/savantskie/persistent-ai-memory.git

---

What's New in v2.0.0 (September 20, 2026)

Full Feature Parity Release -- PAM is now fully in sync with the internal production system. Every feature from months of real-world use is now public.

• Core Identity System -- Distills a complete profile of who the user is across all conversations. Pulls from curated memories, the OpenWebUI memory table, and archived databases. Updates incrementally rather than starting from scratch each time.
• Automated Memory Maintenance -- Background tasks reformat old memories, detect contradictions and updates, and link orphaned memories to their source conversations. Includes a backlog processor that deduplicates entries, fills in missing metadata, and re-ranks memory importance using the LLM.
• Task Coordinator -- Centralized scheduler replaces ad-hoc sleep loops. Proper concurrency control with per-database locking and LLM call gating. Detects user activity and waits until idle to run heavy work.
• Vision Support -- Precomputes image embeddings via a separate vision server so the main LLM does not need mmproj loaded. Caches embeddings for cross-turn injection. Works with llama.cpp, LM Studio, and any OpenAI-compatible embedding provider.
• Cross-System Deduplication -- When promoting short-term memories to long-term storage, checks for exact and semantic duplicates. Merges related memories instead of creating redundant copies.
• Heartbeat-Based Idle Detection -- The system respects when you are actively using it and delays heavy background tasks until you are idle.
• New Utility Tools -- Interactive model-specific memory deletion (pam_delete_model_memories.py), memory export to text files (export_memories.py), and MCP tool call filtering (filter_tool_calls.py).
• Normalization Migration -- One-time migration that normalizes user and model IDs for consistent multi-user isolation.

Upgrade from v1.5.0: Drop-in replacement. Replace your existing PAM files with the new versions. The database schema is backward compatible.

---

What's New in v1.5.0 (March 28, 2026)

Major Architectural Rewrite: OpenWebUI-Native Integration

• ✅ OpenWebUI-first design - AI Memory System now deeply integrated into OpenWebUI via plugin (primary deployment method)
• ✅ Advanced short-term memory - sophisticated memory extraction, filtering, and injection for chat conversations
• ✅ User ID & Model ID isolation - strict multi-tenant support with configurable enforcement for security and tracking
• ✅ Complete system portability - all hardcoded paths replaced with environment variables (works anywhere)
• ✅ Generic class names - removed all Friday-specific branding (FridayMemorySystem → AIMemorySystem)
• ✅ Production-ready - enhanced error handling, validation, and logging throughout

Upgrade from v1.1.0: See CHANGELOG.md for migration guide.

---

📚 Documentation Guide

Choose your starting point: