AI agents have a fundamental problem: they forget everything when the conversation ends. Every new session is a blank slate, forcing users to re-explain context, preferences, and past decisions. EverOS, the open-source memory operating system from EverMind, has been solving this since its v1.0 launch three weeks ago. Now, with EverOS 1.1.0 — released on June 24, 2026 — the project introduces first-class knowledge management, reflection orchestration, and expanded runtime capabilities that bring it closer to a full cognitive architecture for autonomous agents.
Why AI Agents Need a Memory Operating System
Before diving into what's new, it's worth understanding why EverOS exists at all. Modern AI agents — from Claude Code and Codex to OpenClaw and Hermes — are incredibly capable within a single session but lose all context when that session ends. Every new conversation requires rebuilding the agent's understanding of the user, the project, and the work in progress. This statelessness is the single biggest bottleneck holding back autonomous agents from becoming truly useful assistants.
EverOS solves this by providing a persistent, self-evolving memory layer that agents can read from and write to across sessions, platforms, and even different agent tools. It organizes memory into four types borrowed from cognitive science: episodic (what happened), semantic (general knowledge), procedural (how to do things), and profile (who the user is). All memories are stored as plain Markdown files by default — open, version-controllable, and never locked into any proprietary format.
The project has quickly gained traction, amassing over 9,600 stars on GitHub and 810 forks within three weeks of its initial public release. Its Apache 2.0 license, self-hosted deployment option, and compatibility with major agent platforms (including Hermes, Claude Code, Codex, OpenClaw, and MCP) have made it a popular choice for developers building production agent systems.
What's New in EverOS 1.1.0
Version 1.1.0 is the first major feature release since the project's debut, and it significantly expands the memory system beyond basic episode storage. Here are the key additions:
1. Knowledge APIs — Structured Document Management
The headline feature of 1.1.0 is the introduction of Knowledge APIs. While EverOS already handled episodic memory (conversation histories and agent trajectories), the new Knowledge system adds dedicated storage and retrieval for structured documents. Agents can now create, list, patch, delete, and search knowledge documents with full taxonomy support.
What makes this powerful is that knowledge is not just stored — it's indexed with the same hybrid retrieval engine (BM25 + vector search + scalar filtering) that powers EverOS's memory system. An agent can store a project architecture document, a coding style guide, or a client's brand guidelines as a knowledge document, then retrieve it with semantic search during future sessions. The knowledge-topic search feature means agents can even find documents by thematic relevance, not just keyword matches.
For developers building agent-based applications that need long-term reference material — documentation wikis, policy handbooks, technical specs — the Knowledge APIs turn EverOS into a full knowledge management system that's accessible programmatically by any agent in the ecosystem.
2. Reflection Orchestration — Memory That Gets Smarter Over Time
Memory that only accumulates without ever being refined becomes noise. EverOS 1.1.0 introduces Reflection orchestration , a mechanism for periodically merging and refining episode clusters. Instead of storing every raw interaction forever, the system can now consolidate related episodes into distilled summaries, identify patterns across sessions, and remove redundant information.
This is a critical step toward autonomous agent memory management. A support AI that handles 1,000 customer inquiries shouldn't retain all 1,000 raw transcripts — it should distill the patterns, common solutions, and user preferences into compact, reusable knowledge. Reflection orchestration enables exactly that: offline consolidation that transforms raw trajectories into structured, queryable intelligence.
This builds directly on EverOS's existing Self-Evolving Skills pipeline, where successful agent trajectories (Cases) are automatically promoted into reusable Skills through offline distillation. With 1.1.0, the reflection layer adds a second consolidation path focused on episode-level knowledge, not just procedural skills.
3. Expanded OME Runtime
The Open Memory Execution (OME) runtime — the engine that actually runs memory operations — gets several upgrades in 1.1.0. Event IDs give developers traceability for every memory operation, making debugging and auditing easier. The run-record storage migration system allows the runtime to evolve its internal data structures without breaking existing memory stores. Configuration reload behavior has also been improved, so runtime parameters can be updated without full restarts.
These may sound like infrastructure changes, but they matter for production deployments where uptime and observability are critical. A memory system that can't be debugged, upgraded, or reconfigured without downtime isn't ready for production — EverOS 1.1.0's OME improvements directly address these gaps.
4. Better Agent Memory Isolation
The search and retrieval systems for agent-owned memory have been significantly improved. In multi-agent setups — where different agents may need access to different memory scopes — the new owner-type isolation ensures that one agent's skills, cases, and knowledge don't bleed into another's context. This is essential for any production deployment running multiple agents with different roles and access levels.
5. Cleaner Error Handling
EverOS 1.1.0 reworks the entire API error surface with typed exception handlers and consistent error envelopes. Previously, a misconfigured embedding provider might silently return degraded results. Now, missing providers fail explicitly with clear configuration errors. For developers integrating EverOS into their agent workflows, this means faster debugging and fewer mysterious failures.
The Bigger Picture: From Episodic Memory to Full Cognitive Architecture
EverOS's trajectory from v1.0.0 to v1.1.0 reveals a clear roadmap. Version 1.0.0 established the foundation: Markdown-first storage, hybrid retrieval with BM25 and vector search, and three-piece storage using SQLite and LanceDB. Version 1.0.1 added security hardening around caller-supplied identifiers and markdown write containment.
With 1.1.0, EverOS is moving beyond simple memory storage toward a full cognitive architecture for AI agents. Knowledge APIs give agents the ability to manage and query structured information. Reflection transforms raw data into refined knowledge. Together, these features enable agents that don't just remember — they learn, synthesize, and improve over time.
This direction aligns with where the entire AI agent ecosystem is heading. As we noted in our coverage of multi-agent orchestration platforms and Hermes agent hosting, the ability to maintain persistent, evolving context across sessions is the missing piece that turns AI agents from impressive demos into genuinely useful tools.
How to Get Started with EverOS 1.1.0
EverOS is available under the Apache 2.0 license and can be deployed in two ways:
- Self-hosted (OSS) : Clone the GitHub repository, start Docker services, configure environment variables for your LLM and embedding providers, and you're running the full memory stack on your own infrastructure. No external services required — all data stays on your machines.
- EverOS Cloud : A managed, production-ready version available at everos.evermind.ai. The free tier includes 3 memory spaces and 50,000 MCU per month. No credit card needed.
For developers who already use AI coding assistants like Claude Code or Codex, EverOS plugins and integrations are available for all major platforms — including a dedicated Hermes Agent plugin and an MCP server for tool-agnostic access.
Performance That Backs It Up
EverOS isn't just conceptually interesting — it performs well in benchmarks. On the LoCoMo benchmark for long-context memory, EverOS scores 93.05% accuracy , compared to ~45% for naive RAG and ~70% for other memory infrastructure. Retrieval latency sits at under 500ms at p95 , and the system's proprietary retrieval model delivers 7-15x lower token costs compared to naive full-context approaches.
It also supports multimodal ingestion out of the box — PDFs, images, Markdown, Excel files, Word documents, and URLs can all be parsed, chunked, and indexed as memory in a single API call. No separate preprocessing pipelines required.
The Bottom Line
EverOS 1.1.0 is a significant step forward for open-source AI agent memory. The addition of Knowledge APIs and Reflection orchestration transforms it from a simple memory store into a genuine cognitive layer — one that doesn't just record what happened but actively organizes, refines, and surfaces the right information at the right time.
For developers building production agent systems — whether single-agent coding assistants or multi-agent economic platforms — EverOS provides the memory infrastructure that turns stateless tools into persistent collaborators. And with the project moving at this pace (three releases in three weeks), the gap between the v1.1.0 features and the roadmap — Wiki-style knowledge, deeper offline evolution, Dreaming — may close faster than anyone expects.
Links: GitHub Repository | EverOS Homepage | Documentation | Release Notes
Photo by EverMind Team
Originally published on TekMag
Top comments (0)