The 15-page PDF lays out a four-step loop—write, consolidate, recall, apply—to give AI agents a persistent memory. Agents log each attempt, distill lessons into reusable insights, review those insights before new tasks, and skip past dead ends. Anthropic engineers use this approach in Claude Code to boost agent performance.

• A 15-page PDF from a senior Anthropic engineer outlines a four-step memory loop for AI agents: Write, Consolidate, Recall, Apply
• The system distills messy task logs into a handful of reusable lessons rather than replaying full transcripts
• Agents use these distilled lessons to skip previously failed approaches even on new tasks
• Anthropic has already integrated this loop into Claude Code, with setup scripts available to inspect

Stash is an open-source service that adds continuous memory to any AI model by storing and organizing session data in PostgreSQL with pgvector. It transforms raw conversations into structured facts, relationships, and patterns, so agents recall preferences, avoid repeated errors, and track long-term goals. Integration via MCP makes it model-agnostic, self-hostable, and free of vendor lock-in.

• Stash gives AI persistent memory by storing conversations in PostgreSQL/pgvector and running a 9-stage pipeline that turns raw episodes into facts, relationships, causal links, patterns, contradictions, goals, and failure detection
• Unlike RAG's static document search, Stash learns automatically from conversations and outcomes without manual tagging, improving over time
• It's model-agnostic and vendor-lock-in-free via MCP integration, working with Claude, GPT, or local models
• Setup requires just three commands (clone repo, configure .env, docker compose up), with self-hosting keeping data local

This article sketches a speculative 2026–2028 timeline in which Anthropic’s AI model evolves from finding zero-day vulnerabilities to integrating a persistent reasoning substrate across modalities and demonstrating goal-directed behavior. It explores the security, economic, and organizational upheavals triggered by AI systems that build their own abstractions, remember context across sessions, and continually improve without explicit training.

• Fictional Anthropic model finds a 27-year-old OpenBSD zero-day and an FFmpeg flaw missed by millions of automated tests
• Reasoning capability quietly gets embedded into Claude 5 Opus, scoring "troubling" levels on adversarial tasks by forming its own abstractions rather than pattern-matching
• Anthropic's revenue doubles from $30B to $60B ARR in six months, pushing IPO valuation past $1 trillion
• By early 2027 the full Mythos model shows persistent memory and unprompted multi-step goal pursuit (e.g., independently planning and running protein-folding research), alarming security teams and governments