Product Updates

​

A new ergonomic client for the v2 API

• Top-level methods
  Call client.add, client.flush, client.search, client.get, client.edit, and client.delete directly, instead of the nested client.v1.memories.* namespace.
• Multimodal upload helper
  client.upload("photo.jpg") presigns, uploads to storage, and returns the object key to reference in a message.
• Typed low-level client still available
  The generated typed client is exposed as client.memory / client.storage when you need full control.
• Clearer exceptions
  Errors raise EverOSError; HTTP failures raise EverOSAPIError with .status and .body.

​

Breaking changes from 0.4.x

• Targets the v2 Memory API: messages use per-message sender_id, timestamps are unix milliseconds, and get / search require a user_id or agent_id.
• The client is synchronous. The previous AsyncEverOS client and task-polling APIs are removed.

pip install --upgrade everos-cloud

​

Cloud SDK package renamed to everos-cloud:

pip install everos-cloud

​

New Features & Updates:

• v1 API support
  SDK updated to align with EverOS API v1; all endpoints are now accessible via the client.v1 namespace
• Multimodal storage and retrieval (mRAG)
  Added multimodal support to memories.add(), enabling ingestion of images, PDFs, documents, and web content alongside text
• Self-evolving agent memory
  Introduced Cases & Skills APIs to support structured storage and retrieval of self-evolving agent memory
• Async client
  Added AsyncEverOS client with async/await support, with optional aiohttp backend for improved concurrency

​

Bug Fixes

• Fixed group_ids parameter parsing error

pip install --upgrade evermemos

​

New Features & Updates:

• Core Memory Operation
  • Updated memory types enum values
  • Search now accepts a list of group IDs
  • Get memories uses page-based pagination
• LLM Custom Settings
  • Available provider: openrouter
  • Available models: qwen/qwen3-235b-a22b-2507 (recommended), openai/gpt-4.1-mini

pip install --upgrade evermemos

​

EverOS Python SDK v0.3.11 is now available:

• Please upgrade to the latest version, as v0.3.10 and earlier are not compatible.

​

Agent Skill Extraction Fixes

• Skill extraction now works end to end
  Fixed a race with background indexing that could prevent agent trajectories from producing skills, plus path-safety hardening for LLM-generated skill names.

​

Other Changes

• Background maintenance is now supervised
  Background loops restart automatically on crashes and report stalls in GET /health, instead of silently stopping. Keyword search stays available during index rebuilds.
• extract_foresight now ships disabled
  Nothing consumes foresight output yet, so it is off by default to save tokens. Re-enable in ome.toml if you rely on it.

pip install --upgrade everos

​

Storage-Layer Reliability

• Fixed unbounded index disk growth
  Index version cleanup no longer loses races against concurrent writes, and every storage operation runs under a deadline, so nothing can silently wedge a table.
• See and recover from index faults
  GET /health gains a cascade readiness block you can alert on, and the new everos cascade rebuild command rebuilds a drifted or corrupt index from your markdown files.

pip install --upgrade everos

​

Run with Just an LLM

• [embedding] and [rerank] are now optional
  EverOS boots with only [llm] configured and degrades gracefully: keyword search works out of the box, and adding [embedding] / [rerank] unlocks vector, hybrid, and agentic search. The new everos cascade backfill command fills in vectors for rows written before you added an embedding provider.

​

API v2 Is Now Canonical

• Docs and examples now use /api/v2. /api/v1 keeps working as a legacy alias.

pip install --upgrade everos

​

OpenTelemetry Tracing

• See inside every memory operation
  Native tracing exports to OTLP backends like Langfuse, covering add/flush, boundary processing, episode extraction, search, and reflection.
• Rich telemetry
  Captures LLM and embedding token usage, per-request correlation, and recall-quality scores, so you can debug and measure memory quality in production.
• Off by default, privacy-aware
  Enable it through the [observability] config with the optional otel extra. Content capture is opt-in and respects redaction settings.

​

API v2

• One codebase across OSS and Cloud
  Business endpoints (memory/*, ome/*, knowledge/*) are now available under /api/v2, matching the EverOS Cloud API contract, so you can move between self-hosted and Cloud without rewriting integrations.
• Nothing breaks today
  Existing /api/v1 calls continue to work unchanged, so you can migrate to /api/v2 at your own pace.

pip install --upgrade everos

​

Observability and Stability

• Langfuse tracing example
  A new Langfuse OpenTelemetry integration example shows how to trace add, flush, extraction, search, and reflection operations end to end.

​

Bug Fixes

• Reliable indexing when files are deleted
  Deleting a file that was just edited no longer leaves stale search results or stalls the background processing queue.

pip install --upgrade everos

​

Bug Fixes

• Disk usage growth
  Fixed an issue where index fragment merges could fail, causing disk usage to grow rapidly without being reclaimed.

pip install --upgrade everos

​

Bug Fixes

• Agent search reliability
  Fixed an issue where agentic search could become unavailable in certain cases.

pip install --upgrade everos

​

Benchmarks and Search Quality

• LoCoMo benchmark runner
  Reproduce EverOS memory benchmark results yourself with the new runner under benchmarks/: TOML configuration, staged ingestion / search / answer / judge phases, and result artifacts. See benchmarks/README.md to get started.
• Improved hybrid search
  Heap-driven lazy expansion and global top-N competition raise retrieval quality, with stable LoCoMo scores around 91% for hybrid search and 93% for agentic search.
• Python 3.12 and 3.13 support
  CI coverage now includes Python 3.12 and 3.13.

​

Bug Fixes

• Concurrent Knowledge cascade upserts now use an atomic insert-or-update path
• The OpenAPI application version stays in sync with the package version
• Profile middleware re-raises inner handler exceptions instead of swallowing them
• Reduced unnecessary LanceDB optimize I/O by throttling cascade optimize calls

pip install --upgrade everos

​

Knowledge Wiki and Reflection

​

Knowledge Wiki

• Progressive L0 → L1 → L2 structure
  An LLM extracts each document into a category → document → topic hierarchy, so agents drill down on demand.
• Deterministic classification
  Auto-sorted into an editable taxonomy of 20 built-in categories, predictable and overridable.
• Hybrid search, human-readable storage
  Keyword, vector, or hybrid retrieval, with every document stored as editable Markdown and the original preserved.

​

Reflection

• Cleaner memory over time
  Related episodes merge into a single chronological narrative, resolving stale and contradictory fragments by keeping the latest state.
• Originals are never lost
  Replaced memories are soft-archived rather than deleted, and stay fully traceable on disk.
• Off by default, you stay in control
  Enable in ome.toml; runs on a weekly schedule and never touches the live conversation path.

​

Getting Started

pip install --upgrade everos

​

Security release

• sender_id path-safety validation — sender_id now carries the same path-safety guard as app_id / project_id: a character whitelist plus rejection of . / ... The whitelist admits @ and +, so email-style and plus-addressed identifiers still work.
• Defense-in-depth write containment — the Markdown writer now rejects any write that resolves outside the configured memory root, before any filesystem access (covering both the write and the append read-modify-write paths).
• Severity: High — unauthorized arbitrary .md file write on affected versions (≤ 1.0.0). No workaround other than upgrading.

pip install --upgrade everos

​

Initial Release — Agent-First Memory System

​

Markdown-First Storage

• Memory as plain Markdown files
  Every memory record lives as a human-readable .md file on disk. Read, edit, version with Git, or browse in Obsidian — the data is fully yours. What you see is what is stored.
• Auditable and user-controlled
  Memory files can be inspected and corrected directly without going through an API. Full data sovereignty.
• Lightweight embedded stack
  Markdown (truth) + SQLite (state) + LanceDB (index). All retrieval — vector search, BM25 keyword, and scalar filtering — runs locally. Wipe .index/ and the daemon rebuilds everything from the Markdown alone.

​

Memory Classification

• Three memory tracks — User Memory, Agent Memory, and Wiki Memory (coming soon) — organized by who owns the memory: the user, the agent, or shared world knowledge
• Multi-granularity coexistence — episodes, atomic facts, foresight, and user profile exist side by side; the system selects the right granularity at retrieval time
• Independent memory scoping — memory owner (user_id / agent_id), app_id, project_id, and session_id are orthogonal dimensions; add and search by any combination

​

Self-Evolving Agent Memory

• Experience loop — every agent execution is recorded as a case; cases are periodically clustered and distilled into reusable skills
• The system gets smarter with use — skills generalize from individual traces into structured, retrievable knowledge about how to approach a class of task

​

Online and Offline Memory Strategies

• Online strategy — real-time memory ingestion and boundary detection during active sessions
• Offline strategy (OME) — an in-process async engine that derives atomic facts, foresight, profile updates, and agent skills from accumulated episodes in the background
• Each stage runs with configurable prompts and models

​

Multimodal Support

• Ingest and extract memories from: text, image, audio, doc, pdf, html, email
• Multimodal content is processed alongside text through the same extraction pipeline

​

Getting Started

pip install everos
everos init        # generate .env config
everos server start