Nexus Memory

Nexus Memory gives Claude Code persistent memory across sessions using a Qdrant instance. The same memory store can be shared with Hermes Agent and OpenClaw - one brain, many agents.

How it works

• Auto-Recall (UserPromptSubmit hook): Before each prompt, searches Qdrant for relevant memories and injects them as context.
• Auto-Capture (Stop hook): After each turn, extracts notable facts from the transcript and stores them in Qdrant.
• Session Start (SessionStart hook): Loads project-related memories when a session begins or resumes.

Read this before you install: what gets stored, and where it goes

This skill stores to persistent memory automatically. Facts are extracted from your conversations after every turn and written to Qdrant, then re-injected into later prompts without asking again. Two things deserve a deliberate decision:

1. Conversation content is persisted. Anything said in a session can end up as a stored memory, including details you did not intend to keep. Do not run this around credentials, secrets, regulated data, or client material you are not permitted to retain - unless you have decided how to handle review, deletion and disabling (see "Disabling and controlling it"). Stored memories also resurface later, where they can be mistaken for something just said.

2. Where the text goes depends on your embedding provider - and the default is local. Out of the box the plugin embeds through the local Nexus service (nexus-memory serve), which runs the engine with its local HuggingFace model: nothing has to be installed and memory text does not leave your machine. Only if you deliberately name a cloud provider (Voyage, OpenAI, Google, Jina) is the text of your memories sent to that provider's API. If you switch providers later, re-embed, because vectors from different models are not comparable.

3. A shared collection is readable across agents. If Claude Code, Hermes Agent and OpenClaw point at the same Qdrant collection (the common setup), a memory written by one agent is visible to the others. That crosses tool and trust boundaries - use a separate NEXUS_COLLECTION per project when contexts must not mix.

Disabling and controlling it

• Turn off automatic storage: remove or disable the Stop hook in the plugin's hook configuration (hooks/nexus-hooks.json, or Claude Code's hooks settings) - that hook is what captures facts after each turn. The manual remember / recall / forget tools keep working without it.
• Keep memory local: this is already the default (the local Nexus service, on your machine) - just do not set a cloud provider.
• Keep memory separate: set a distinct NEXUS_COLLECTION per project.
• Remove something: use the forget tool on the memory you do not want.

Configuration

Environment variables (set in .env or shell):

Embeddings

The plugin defaults to the local Nexus service - no API key, no account, nothing leaving your machine, and nothing to install:

Setup (once): run the service - the wizard/installer already registers it:

nexus-memory serve        # http://127.0.0.1:9122, installed as a user service

Point the plugin elsewhere with NEXUS_SERVE_URL if the service listens somewhere else. If it is not reachable, the hooks say so on stderr and skip their work instead of storing or recalling nothing in silence.

The explicit Ollama path is the whole setup then: qwen3-embedding:0.6b is multilingual and instruction-aware, which matters if your memory is not English-only.

bge-m3 (1.2 GB, 1024d, multilingual) is a heavier local alternative. Use one model consistently - mixing models inside a single collection produces incomparable vectors.

Manual Tools

The MCP server (nexus-memory) provides explicit tools:

• remember - Store a memory
• recall - Search memories
• forget - Delete a memory

Shared Store

Same Qdrant collection as:

• Hermes Agent (native plugin)
• OpenClaw (native plugin)
• Any MCP-compatible agent

A memory stored by Claude Code is immediately visible to Hermes and vice versa - one brain, many agents, as long as you want them to share.