.framework-timeline
Comparing the independent development of the NOVA framework with parallel research and industry developments in modular AI memory, agent-first development, and structured context management.
Multi-tier memory architecture with LLM as virtual memory manager, paging information in/out of limited context window.
Even modest ~2k-token contexts cause models to miss facts in the middle. Simply increasing window size is insufficient — retrieval + structure required.
Hierarchical memory organization via segment-level recurrence. 2–57× fewer parameters, 2.5–116× less inference memory vs long-context baselines.
Tree-structured memory for long-form dialogue. Conditional tree traversal instead of linear history or flat vectors.
Light long-range model builds global memory of long context. Memory as learned intermediate structure instead of raw chunks.
Structured, graph-based external memory with explicit edit operations and RL-based memory selection. Deployed in real smartphone assistant.
Open protocol for connecting AI assistants to tools and data sources. LSP-inspired design solving the N×M integration problem.
Multi-agent framework with asynchronous messaging, pluggable memory, and event-driven runtime with learning capabilities.
Practical limits hit with IDE-extended context for AI coding assistants. Even with bigger windows, multi-file workflows don't scale. Begin experimenting with segmented conversation threads (proto-shards).
Built external memory system using modular units called 'shards' — atomic thought-packets with metadata, tagging, and relevance scoring.
Formal whitepaper defining NOVA as a user-centric framework for modular cognitive augmentation through stateless AI systems. Grounded in Extended Mind Thesis, Distributed Cognition, and Working Memory theory.
ADHD-focused productivity assistant using on-device ML with human-in-the-loop design. Parallel thinking about AI as co-regulator inside structured environment.
Zettelkasten-inspired memory: structured notes with attributes, tags, inter-note links. Almost exact parallel to NOVA's shard abstraction.
Recursive problem decomposition with self-verification. Large gains on math tasks without parameter scaling.
First working implementation released as FastAPI server with REST API for shard interaction, creation, semantic search, and listing. Open-sourced with full documentation suite including whitepaper, executive summary, shard memory architecture, and unified consciousness model.
Formal mapping of cognitive functions to shard operations. Defines working memory as active loaded shards, attention as user-selected shards, long-term memory as shard index + embeddings, executive function as user-led shard management, and metacognition as cross-shard synthesis.
Agent-first SDLC: autonomous agents continuously write, test, deploy code. Humans become intent designers and curators.
Microsoft made the Semantic Kernel Agent framework generally available in SK 1.45 (.NET) and 1.27 (Python), providing a production-ready SDK for building AI agents and multi-agent systems with improved memory and orchestration primitives.
OpenAI updated ChatGPT memory so that, beyond explicit saved memories, it now references all past conversations via chat history to deliver more personalized, cross-session responses, with granular user controls for both channels.
Rebuilt NOVA as MCP-compliant server with 7 tools and 2 resources, removing OpenAI dependency from the server layer. Cleaner separation: server manages shards, connected LLM handles reasoning.
OpenAI integrates MCP across products. Windows 11 marketed as 'agentic OS' with native MCP support. Protocol transitions from experiment to default integration layer.
OpenAI launched Codex as a cloud-based software engineering agent in ChatGPT, powered by the codex-1 model, able to read repositories, run tests, and autonomously implement features and bug fixes in sandboxed environments.
Microsoft announced the general availability of Azure AI Foundry Agent Service, enabling orchestration of multiple specialized agents with Agent-to-Agent (A2A) and Model Context Protocol support, unifying Semantic Kernel and AutoGen into a single agent-focused SDK.
Google introduced Jules as an asynchronous agentic coding assistant that clones repositories into secure cloud VMs, autonomously performs tasks like writing tests, fixing bugs, and adding features, then returns plans, reasoning, and diffs integrated into GitHub workflows.
Mistral announced its Agents API, a framework for building autonomous agents with built-in connectors for Python execution, web search, image generation, document retrieval, MCP tools, persistent conversational memory, and orchestration of collaborating agents.
OpenAI expanded its Codex software engineering agent and Agents SDK, adding broader access, TypeScript support, human-in-the-loop approvals, serialized agent state, and improved tracing.
Anthropic detailed Claude's Research feature as a multi-agent system using an orchestrator-worker pattern where a lead agent plans research, spawns specialized subagents that search in parallel, and aggregates their findings.
Industry analysis of trade-offs between monolithic and modular AI systems. NOVA is explicitly modular cognitive architecture.
Google announced Agent Mode for Gemini in Android Studio, an experimental agentic capability that uses built-in IDE tools so Gemini can autonomously assist with Android app development tasks.
Mistral introduced Devstral Medium and an upgraded Devstral Small, coding models optimized for agentic workloads and generalized to different prompts and scaffolds.
AI agent does implementation, docs, tests, versioning. Human 'Editor' sets goals, constraints, acceptance criteria.
Claude Code running as background agents in remote dev environments. Multiple agents working asynchronously, persisting state.
LangChain published its Deep Agents architecture and released the deepagents Python package, defining agents that combine planning tools, subagents, filesystem-based persistent memory, and detailed prompts to handle long-horizon multi-step tasks.
Survey of persistent memory patterns: episodic tables, hierarchical stores, multi-agent shared memory, selective retention.
Google moved Jules out of beta into general availability, powered by Gemini 2.5 Pro, adding higher-limit Pro and Ultra tiers for intensive multi-agent coding workflows.
xAI released grok-code-fast-1, a dedicated agentic coding model built with programming-focused pretraining and post-training on real pull requests, optimized for terminal and file-editing tools.
LangChain announced 1.0 alpha versions of LangChain and LangGraph, repositioning LangGraph as a low-level durable agent orchestration runtime and LangChain as an agent-centric framework built on top.
Three-phase evolution toward agentic AI. Planning, tool use, reflection, collaboration, and memory as core design patterns.
Anthropic published the Claude Agent SDK, a general-purpose agent harness that exposes Claude's agent loop, tools, and context management in Python and TypeScript for building autonomous agents.
Google introduced Jules Tools, a command-line companion that makes the Jules asynchronous coding agent programmable and scriptable from the terminal.
Google Cloud launched Gemini Enterprise, a unified agentic AI platform combining Gemini models, first- and third-party agents, and orchestration technology to build multi-step data-connected enterprise agents.
Microsoft announced the public preview of Microsoft Agent Framework, an open-source unified SDK and runtime merging AutoGen's multi-agent orchestration with Semantic Kernel foundations.
Anthropic introduced Agent Skills, filesystem-based folders of instructions, scripts, and resources that Claude can dynamically load to specialize on tasks across Claude apps, Claude Code, and the API.
LangChain released stable versions of LangChain 1.0 and LangGraph 1.0, the latter providing durable agent runtime with graph-based execution, checkpointing, short-term memory, and human-in-the-loop.
LangChain shipped deepagents v0.2, adding pluggable storage backends, large tool-result eviction, conversation-history summarization, and repair of interrupted tool calls.
Released a modular, model-agnostic toolkit refactoring the OpenHands agent stack into a composable SDK with event-sourced state, pluggable tools, multi-LLM routing, and sandboxed execution.
Research paper introducing a software engineering agent that starts from a minimal scaffold and autonomously evolves its own tools and prompts at runtime.
Google launched Antigravity, an agentic development platform pairing an AI-powered editor with a Manager Surface for spawning and orchestrating multiple coding agents that learn from a shared knowledge base.
Anthropic described engineering patterns in the Claude Agent SDK for long-running agents, including initializer and coding agents plus compaction strategies for multi-context-window work.
Per-user memory storing prior interactions. Historical context for future coding sessions — persistent, user-scoped memory in production.
MCP donated to Agentic AI Foundation under Linux Foundation. Major spec release introducing Tasks for long-running operations.
Research paper presenting a fully autonomous multi-agent coding framework combining blueprint distillation, stateful code memory, retrieval-augmented knowledge injection, and closed-loop error correction.
Mistral released Devstral 2 and the open-source Mistral Vibe CLI, a terminal-native coding agent for repository-scale agentic coding tasks.
Research introducing CCA and the Confucius SDK, an open agent development platform with hierarchical working memory, persistent note-taking, modular tools, and a meta-agent that designs and refines coding agents.
Google added the Interactions API in beta as a unified interface for interacting with Gemini models and agents, and launched the Gemini Deep Research Agent for autonomous multi-step research tasks.
xAI launched the Grok Voice Agent API, compatible with the OpenAI Realtime API, for building multilingual voice agents with tool calling and real-time data search.
Stateful, server-side agent runtime with persistent memory, background execution, and integrated tool calls. Positioned as 'remote operating system' for agents.
Anthropic updated Agent Skills to be an open standard, specifying portable filesystem-based skill packages that work across Claude apps, Claude Code, the Agent SDK, and third-party runtimes.
Defines Continuum Memory Architecture (CMA): persistent, selectively retained, temporally chained, abstracted memories. Critiques classic RAG as 'stateless lookup'.
Anthropic announced that Apple's Xcode now supports the Claude Agent SDK, enabling Claude-powered agents to handle long-running coding tasks within the Xcode IDE.
Google DeepMind proposed a framework for intelligent AI delegation on the agentic web, defining pillars like dynamic assessment, adaptive coordination, verifiable completion, trust and reputation, and security.
Microsoft Agent Framework reached Release Candidate status for .NET and Python, finalizing a stable API for single and multi-agent systems with checkpointing, human-in-the-loop, and multi-provider model support.
Codex App Server as standardized harness. JSON-RPC protocol with items, turns, threads. Persistent agent sessions with shared state.
Every line of code written by Codex agents — ~10× speedup. The crucial ingredient is not the model but the harness. 'Humans steer. Agents execute.'
CodePath redesigns curriculum with Claude and Claude Code as central. AI-assisted development as default baseline.
MuleRun launched a self-evolving personal AI platform giving each user a dedicated 24/7 cloud VM with a three-tier evolution engine covering task memory, domain skill acquisition, and community knowledge sharing.
LangChain published a detailed post on context engineering for deep agents, covering persistent memory, skill loading, and long-horizon task management.
Google published a blog post describing the problem of agents lacking specialized knowledge and how skill files loaded at invocation time address it.
Hidden features surfaced in Gemini for Business builds showing a Skills tab with pre-made skills for code review and PRD writing, plus a Skill Architect meta-skill for creating custom skills with name, description, and instructions fields.
xAI is developing a Custom Skills feature for Grok, letting users create and import modular instruction sets, currently visible in code hints but not yet launched.
External, structured, modular memory with orchestration layers is the emerging architectural consensus for AI agent cognition.