Agent & Memory Architectures: Skills, Plan Mode, Hindsight, Titans

New agent and memory architectures emphasize modular skills, plan-first IDE modes, reflective memory, and long-term recall.
Key insights
Skills replace blind tool‑search in agent design
Authors argue that the 'tool search' pattern—dumping many function definitions into a prompt—fails at scale. The proposed replacement is a 'skills' paradigm that packages tools with usage guidance, prompts, and best practices, and is loaded on demand via a classifier or router. Skills reduce context bloat, improve reliability, and provide implicit 'how/when' usage knowledge that raw API signatures lack. The advocated flow uses intent classification to select and inject only a few relevant skills into the model context, improving execution reliability and scalability.
Plan Mode in AI IDEs separates planning from execution
Armin Ronacher and others describe 'Plan Mode' as an AI IDE pattern where the model first generates a detailed implementation plan—files to read/modify, steps, and required contexts—before performing edits. Native Plan Mode tightly integrates with the IDE, supports permissioned step execution, and enables state rollback and user pre‑approval of steps. This decoupling reduces logic errors, preserves intent across sessions, and shifts human roles from micromanaging edits to architectural review.
Hindsight proposes structured memory with reflect-and-update capabilities
Hindsight is an open memory architecture that organizes memory into four networks—world facts, agent experience, entity summaries, and evolving beliefs—and implements three operations: retain, recall, and reflect. Reflection actively reasons over memory to update beliefs and correct errors. Papers and open-source projects show Hindsight substantially improves long-term agent performance on LongMemEval and similar tasks, reducing hallucinations through structured memory and active revision.
Titans architecture achieves human-like memory dynamics
Summaries of Titans research describe a tri-level memory approach combining short-term attention, long-term parameterized memory, and persistent task memory. Titans uses surprise-based retention, adaptive forgetting, and test-time memory updates to learn during inference. Reported results include high accuracy on long-context tasks (e.g., 96% on needle-in-haystack tasks at 16k tokens) and strong performance on language modeling, time series, and genomics, often with far fewer parameters than competitors.
Prompt caching reduces cost and latency by reusing KV states
Prompt Caching is explained as caching KV matrices (not raw text) to avoid recomputing attention, yielding up to ~90% token cost reduction and ~85% latency reduction for long-context scenarios. Implementations differ: OpenAI's automatic routing yields variable hit rates (~50%), while Anthropic's explicit cache breakpoints yield stable hits (~100%) in structured setups. Cache lifetimes tend to be short (5–10 minutes) unless refreshed by frequent requests.