EverOS OpenClaw Plugin

Persistent, cross-session memory for OpenClaw, backed by a self-hosted EverOS — through natural conversation.

This plugin claims OpenClaw's exclusive memory slot and wires the OpenClaw lifecycle to a local EverOS server (/api/v2/memory/* on 127.0.0.1:8000).

v3.0.0 is a major-version replacement. Versions ≤ 2.x were context-engine plugins targeting the old EverMemOS API (/api/v1/memories/* at :1995). This version is the memory-slot plugin for the current EverOS (/api/v2/memory/* at :8000). Requires EverOS ≥ 1.2.3 and OpenClaw ≥ 2026.8.1 (2.0). The package declares this host/API floor in OpenClaw's install-time compatibility metadata, so an older Gateway is rejected before plugin code loads.

What it does

• Recalls relevant memories before every reply and injects them as context
• Saves every finished turn after it ends — text, staged images/audio/documents, and full tool-call trajectories
• Seals the conversation tail when a session ends (/new, /reset, shutdown — including clients whose /new never notifies the gateway)
• Auto-starts a local EverOS if one isn't already running (detect-then-provision)
• You just chat — no memory_store / memory_search tool calls, ever

Good to know:

• This is a memory-slot plugin — installing it displaces the stock memory-core (the installer switches the slot for you)
• Fail-open by design: if EverOS is down or unreachable, OpenClaw keeps working normally — memory just pauses
• Zero runtime dependencies (native fetch)

Quick start

Recommended install:

npx --yes --package @everos-ai/openclaw-plugin everos-setup

The installer will:

• ask before accepting the plugin's declared memory capability, then run the official openclaw plugins install (claims the exclusive memory slot)
• ask before granting conversation access — required for both recall and saving, never granted silently
• help point the plugin at your EverOS checkout when everos isn't on the gateway's PATH
• restart the gateway and health-check the result

Non-interactive / scripted installs: everos-setup --grant --everos-dir /path/to/EverOS (--grant also accepts the declared memory capability). To install while leaving recall/capture off, use --accept-capabilities --no-grant. All flags are listed by everos-setup --help. Prefer doing it by hand? See Manual install.

Then verify with natural language — just mention something about yourself:

My favorite coffee is espresso.

then start a new session (/new), send any message, and ask:

What coffee do I like?

(Give it a few seconds between turns — extraction runs asynchronously in EverOS.)

Do not set hooks.allowPromptInjection to false — it defaults to on, and recall needs it to inject memory into the prompt.

Backend

Default backend address:

http://127.0.0.1:8000

Health check:

curl http://127.0.0.1:8000/health

If EverOS is already running, the plugin detects and uses it. If it isn't, the plugin starts one itself (everos server start, forced into agent mode on the configured port). When EverOS lives in a project virtualenv, everos-setup wires the start command for you (it prompts for your checkout and sets EVEROS_OC_START_CMD / EVEROS_OC_EVEROS_DIR — see Config reference and Troubleshooting if you ever need to adjust them by hand).

Setting up EverOS from scratch:

git clone https://github.com/EverMind-AI/EverOS.git
cd EverOS
uv sync
uv run everos init      # creates ~/.everos/everos.toml (+ ome.toml) — REQUIRED before first start
# edit ~/.everos/everos.toml — fill in the api_key fields (LLM / embedding / rerank)
uv run everos server start

Images, audio, documents, and video

The plugin consumes OpenClaw 2.0's canonical, staged inbound media facts and forwards supported assets to EverOS. Enable EverOS's optional parser in the EverOS checkout:

uv sync --extra multimodal

Then configure the parser's separate provider in ~/.everos/everos.toml. This is not the main [llm] key:

[multimodal]
model = "google/gemini-3-flash-preview"
base_url = "https://openrouter.ai/api/v1"
api_key = "<your key>"

The selected endpoint/model must accept image_url parts and, for voice notes, audio parts. Restart EverOS and verify that /health reports capabilities.multimodal_llm: true and does not list multimodal_upload under disabled_features.

• Supported end to end: images, audio, PDF, HTML, email, and office documents. Office files additionally require LibreOffice on the EverOS host.
• Staged local files are sent as file:// URIs, so OpenClaw and EverOS must see the same filesystem path. If EverOS runs elsewhere, use an HTTP(S) media URL. Restrict readable local paths with [multimodal].file_uri_allow_dirs when the service is exposed beyond loopback.
• EverOS 1.2.3 has no raw video ContentItem type. If OpenClaw's own media understanding produces a transcript/description for an inbound video, the plugin saves that resulting text, but deliberately does not send the raw video file as a fake audio/image item.
• If multimodal parsing is unavailable, the plugin retries the turn text-only after a definite 415, 422, or CAPABILITY_UNAVAILABLE response.

How natural-language memory works

1. You send a normal message.
2. before_prompt_build — the plugin searches EverOS (developer track + agent track) with a query built from your prompt.
3. Hits are injected as a clearly-fenced block of untrusted historical context — recalled memory informs the model, it can't issue instructions.
4. OpenClaw replies normally.
5. message_received + agent_end — the whole turn (user/assistant text, tool calls/results, and staged images/audio/documents) is forwarded to EverOS /add.
6. EverOS extracts memory on topic boundaries as you chat; when a session ends — /new, /reset, gateway shutdown, or a client-side session switch — the plugin flushes the buffered tail so the last topic is never lost.

So the day-to-day experience is just conversation:

Today: "I prefer dark mode, by the way." Days later, brand-new session: "What UI style do I prefer?" → "Dark mode."

You never run a save command and never search anything — mentioning something is enough for it to be remembered, and asking is enough to recall it. No "remember this" prefix required.

OpenClaw config example

The installed shape in ~/.openclaw/openclaw.json:

{
  "plugins": {
    "slots": {
      "memory": "evermind-ai-everos"
    },
    "entries": {
      "evermind-ai-everos": {
        "enabled": true,
        "hooks": {
          "allowConversationAccess": true
        },
        "config": {
          "EVEROS_OC_BASE_URL": "http://127.0.0.1:8000",
          "EVEROS_OC_USER_ID": "your-name",
          "EVEROS_OC_AGENT_ID": "openclaw",
          "EVEROS_OC_QUERY_N": 1,
          "EVEROS_OC_QUERY_MAX_CHARS": 500
        }
      }
    }
  }
}

Config reference

Every key can be set two equivalent ways: env vars on the gateway process (take precedence) or host-managed plugin config (openclaw config set plugins.entries.evermind-ai-everos.config.<VAR> <value>). Empty/blank values are treated as unset.

Var	Default	Purpose
EVEROS_OC_BASE_URL	http://127.0.0.1:8000	Where EverOS is. Scheme-less values (localhost:8000) are normalized to http://
EVEROS_OC_USER_ID	$USER / $USERNAME / OS account	Developer identity for the user memory track
EVEROS_OC_AGENT_ID	openclaw	Constant pooled agent identity for the agent track
EVEROS_OC_QUERY_N	1	Recent user messages blended into the recall query
EVEROS_OC_QUERY_MAX_CHARS	500	Recall query head-clip budget (chars)
EVEROS_OC_START_CMD	everos server start	Command to auto-start EverOS. Quotes group a path with spaces
EVEROS_OC_EVEROS_DIR	(gateway cwd)	Working directory for the auto-started EverOS

What gets recalled and captured

Recalled

Recall injects up to four sections, all served by EverOS:

• Developer profile — durable facts and preferences about you
• Relevant past episodes — summarized prior conversations
• Relevant cases — concrete past agent trajectories (what worked)
• Relevant skills — reusable patterns distilled from multiple cases

Captured

• User and assistant text (the plugin strips its own injected recall block first, so memory never re-ingests itself)
• Assistant tool calls and tool results, chained by tool_call_id
• Images, audio, PDF, HTML, email, and office documents (inline payload or staged URI). If EverOS definitively rejects multimodal input, the turn retries text-only so its conversation text is not lost
• Video-derived text produced by OpenClaw; raw video bytes are not sent because EverOS 1.2.3 does not define a video content type
• Oversized turns are chunked to EverOS's 500-message limit, in order

Manual install

openclaw plugins install @everos-ai/openclaw-plugin --accept-capabilities

Then grant conversation access — required, one time. OpenClaw 2.0 blocks both before_prompt_build recall and agent_end capture for non-bundled plugins without this grant:

openclaw config set 'plugins.entries.evermind-ai-everos.hooks.allowConversationAccess' true
openclaw gateway restart

If everos lives in a project virtualenv, also set EVEROS_OC_START_CMD / EVEROS_OC_EVEROS_DIR (see Config reference) — or just run everos-setup, which wires them for you.

Troubleshooting

Problem	Fix
No recall and nothing is saved	Grant conversation access: openclaw config set 'plugins.entries.evermind-ai-everos.hooks.allowConversationAccess' true, then restart the gateway. The plugin logs a warning when it detects this state.
Saves work but recall does not	Remove an explicit hooks.allowPromptInjection: false override (or set it to true), then restart the gateway.
Audio/image/document becomes a placeholder	Install EverOS with the multimodal extra, configure [multimodal], restart EverOS, and inspect /health. For a local file, also ensure the EverOS process can read that staged path.
Backend connection failed	Check EVEROS_OC_BASE_URL, then curl <baseUrl>/health
Auto-start never brings EverOS up	The gateway can't find everos — set EVEROS_OC_START_CMD to the absolute binary path and EVEROS_OC_EVEROS_DIR to the EverOS repo. Also check nothing else holds the single-instance lock (~/.everos/.index/sqlite/ome.db.lock)
Asked right after telling — no memory yet	Extraction is asynchronous; wait a few seconds. Mid-conversation extraction triggers on topic changes; session end seals the rest
"user-track memory is DISABLED" warning	No user id could be resolved — set EVEROS_OC_USER_ID
Conflicts with another memory plugin	This plugin owns the exclusive memory slot; check plugins.slots.memory is evermind-ai-everos

Files

• dist/index.js — plugin entry (openclaw.extensions)
• src/setup.ts / src/setup-cli.ts — the everos-setup one-command installer
• src/register.ts — slot claim, hook wiring, provisioning service
• src/handlers.ts — staged media / recall / capture / flush (+ session-switch safety net)
• src/everos.ts — typed EverOS REST client (/add, /search, /flush, /health)
• src/provision.ts — detect-then-provision of the EverOS server
• src/config.ts — EVEROS_OC_* configuration
• openclaw.plugin.json — plugin manifest + config schema

Develop

npm install
npm run build        # tsc → dist/
npm test             # unit tests; live smoke runs only if EverOS is up
npm run ci           # lint + local/real-OpenClaw-2.0 typecheck + build + test

License

Apache-2.0