Memory
Why Your AI Needs a Memory System
March 2026
LLMs are stateless by design. Every request starts fresh with no recollection of past interactions. Dream-Weaver changes this by adding persistent, semantically-searchable memory that spans sessions.
Working memory handles the current conversation. Episodic memory stores and retrieves past interactions by semantic similarity. And the promotion pipeline automatically moves frequently-accessed context from working to long-term storage.
The result: your AI assistant remembers what you told it last week, knows your preferences, and builds on prior conversations instead of starting from scratch.
Reasoning
Symbolic Reasoning Meets LLMs: The LSR Integration
March 2026
LLMs are great at language but bad at logic. They hallucinate facts, miss contradictions, and can't prove anything formally. Dream-Weaver's LSR pipeline adds a symbolic reasoning layer that runs automatically on every request.
Nine stages fire in sequence: fact extraction, compression, temporal decay, contradiction detection, validation, ethics gating, working memory with ACT-R activation dynamics, context optimization, and analogical reasoning. When inference fails, the system explains exactly what's missing and suggests fixes.
The symbolic layer doesn't replace the LLM — it augments it. Facts are checked before they reach the model. Contradictions are flagged. And ILP rule learning discovers new logical rules from positive and negative examples.
Routing
Cost-Aware Routing: How We Cut API Costs by 40%
March 2026
Not every query needs GPT-4. Dream-Weaver's Mixture-of-Experts router classifies requests by category — coding, reasoning, creative, conversation, vision, function calling — and routes them to the most cost-effective model that can handle the task.
Simple questions go to local Ollama. Complex reasoning goes to Claude. Code generation goes to specialized models. With circuit breakers for automatic failover and per-tenant quotas for spend control, teams can set budgets and let the router optimize within them.
In production with Pal (our own AI assistant), this approach reduced API costs by 40% while maintaining response quality — measured by user satisfaction and task completion rates.