Hermes

community

The integration contains a standard Hermes MemoryProvider and a standalone slash-command plugin. Hermes remains responsible for the conversation and memory lifecycle; the provider sends recall, capture, and explicit operations to a separately running PowerContext Server. The companion registers /pc and /powercontext before the provider is activated. Backend failures do not interrupt the Hermes conversation.

Prerequisites

Hermes Agent 0.20.4 or newer, available on PATH;
PowerContext CLI and Server installed from master;
a running PowerContext Server.

uv tool install --force "powercontext[cli,server] @ git+https://github.com/oceanbase/powercontext.git@master"
powercontext server run

Install both plugins

In another terminal, install or refresh both plugins from the matching revision:

powercontext setup hermes --source oceanbase/powercontext --ref master
powercontext doctor hermes

The setup command copies the provider to $HERMES_HOME/plugins/powercontext, installs $HERMES_HOME/plugins/powercontext-command, and enables the companion without granting built-in tool override permissions. It does not start the Server or select the provider in Hermes.

Run the Hermes memory setup wizard and select PowerContext:

hermes memory setup

On Hermes 0.20.4, use the generic command above. hermes memory setup powercontext selects the provider but does not open its configuration wizard. Restart Hermes after setup.

Verify recall and writes

hermes powercontext status
hermes powercontext remember preference "The user prefers uv"
hermes powercontext search "Python package manager"

Inside an interactive Hermes session, /pc status should reach the same active provider. Use /pc followed by Tab/Down to inspect the available Memory, Handoff, Experience, Skill, review, statistics, trace, and Scope commands. Hermes 0.20.4 does not provide enough invocation context to route gateway slash commands safely, so the companion rejects gateway invocations; use the provider's Hermes tools in gateway sessions.

The provider uses http://127.0.0.1:8000 by default. The Server resolves an explicit Scope first, then durable session and workspace bindings, and finally its default Scope. Hermes hashes the workspace path only as an external binding key; it does not generate Scope IDs from profiles, users, repositories, or directories.

Configure the connection

The wizard writes non-sensitive settings to $HERMES_HOME/powercontext/config.json. Environment variables override the file:

Variable	Purpose
`POWERCONTEXT_HERMES_CONFIG`	Config file path; defaults to `$HERMES_HOME/powercontext/config.json`
`POWERCONTEXT_HERMES_BASE_URL`	PowerContext Server URL
`POWERCONTEXT_HERMES_AUTHORIZATION`	Complete authorization header, such as `Bearer <token>`
`POWERCONTEXT_HERMES_TOKEN`	Bare-token shorthand used when `AUTHORIZATION` is absent
`POWERCONTEXT_HERMES_SCOPE_ID`	Explicit server-owned Scope ID
`POWERCONTEXT_HERMES_MAX_BYTES`	Maximum prepared-context size, from 512 to 32768 bytes
`POWERCONTEXT_HERMES_TIMEOUT`	HTTP request timeout in seconds
`POWERCONTEXT_HERMES_CAPTURE_TURNS`	Capture completed turns as Sources
`POWERCONTEXT_HERMES_FLUSH_ON_SESSION_END`	Run Memory extraction at session end
`POWERCONTEXT_HERMES_CAPTURE_PRE_COMPRESS`	Capture filtered new turns before compression; disabled by default
`POWERCONTEXT_HERMES_EVALUATION_TRACE`	Record recalled context in sensitive local JSONL traces; disabled by default
`POWERCONTEXT_HERMES_EVALUATION_TRACE_PATH`	Override the evaluation trace directory

Let the Hermes wizard store authorization in its protected .env secret store; do not put the token in config.json. Use plain HTTP only for a loopback Server. See Deploy the Server before connecting to a remote deployment. Evaluation traces contain prompts and recalled context; keep them local and protect them as sensitive data.

Enable automatic extraction only when needed

Completed-turn capture creates Source evidence. It does not create Memory by itself. Automatic Source-to-Memory extraction requires a generation model on the Server:

export POWERCONTEXT_SERVER_INFERENCE_GENERATION_MODEL=provider:model-name
powercontext server run
powercontext capabilities

The capability output must report Memory extraction as enabled. Explicit hermes powercontext remember writes do not require a model.