memory-infino

An OpenClaw memory plugin backed by Infino — hybrid (BM25 + vector) recall on object storage, in one engine.

Status: early. Implements OpenClaw's memory-plugin contract — the three tools (memory_recall / memory_store / memory_forget) plus auto-recall (relevant memories injected into context each turn) and auto-capture (salient user statements stored automatically). Built on the published @infino-ai/infino Node binding.

Why infino for agent memory

• Hybrid recall in one engine — memory_recall fuses BM25 + vector with Reciprocal Rank Fusion. No separate keyword index, no rerank service.
• Seamless by default — auto-recall injects relevant memories into context each turn and auto-capture stores salient statements automatically; the explicit tools stay for manual control. Both are configurable.
• Object-storage-native — memory lives on a local path or S3/Azure; no daemon.
• SQL over memory — the underlying store also supports structured/temporal queries (used internally; a future tool surface).

Tools

Tool	What it does
memory_recall	Hybrid (BM25 + vector) search over long-term memory
memory_store	Save a memory (embedded on write)
memory_forget	Delete by id, or find-and-delete by query

Install

Install the plugin into OpenClaw and select it:

openclaw plugins install @infino-ai/memory-infino
openclaw plugins enable memory-infino          # plugin id is "memory-infino"
# set plugins.slots.memory = "memory-infino" + an embedder (see Configure)
openclaw                                        # restart the gateway

Note the two names: the npm package is @infino-ai/memory-infino; the OpenClaw plugin id (used for plugins.slots.memory and plugins.entries) is memory-infino.

Configure

Select the plugin for the memory slot and give it an embedder + a path:

{
  plugins: {
    slots: { memory: "memory-infino" },
    entries: {
      "memory-infino": {
        enabled: true,
        config: {
          // Default: local in-process embedder — no API key, no service. all-MiniLM-L6-v2 = 384-dim.
          embedding: { provider: "local", model: "Xenova/all-MiniLM-L6-v2", dimensions: 384 },
          // or Ollama: { provider: "ollama", baseUrl: "http://127.0.0.1:11434", model: "mxbai-embed-large", dimensions: 1024 }
          // or OpenAI: { provider: "openai", apiKey: "${OPENAI_API_KEY}", model: "text-embedding-3-small" }
          dbPath: "memory/infino",          // local path, or s3://… / az://… for object storage
          // storageOptions: { aws_access_key_id: "${AWS_ACCESS_KEY_ID}", aws_secret_access_key: "${AWS_SECRET_ACCESS_KEY}", aws_region: "us-east-1" }
        },
      },
    },
  },
}

Key	Meaning
embedding	provider/model/dimensions (+ apiKey/baseUrl). The plugin embeds text via this provider (infino is bring-your-own-vectors), so dimensions must match the model (infino requires 16–4096). Default provider is local.
dbPath	local path or s3:///az:// URI (default memory/infino)
nCent	infino IVF centroid count; 1 = exact (right for memory-scale stores)
compactEvery	auto-merge accumulated superfiles after this many stores (each store is one commit); keeps recall fast as memory grows. 0 disables. Default 128.
autoRecall	inject relevant memories into context automatically each turn (no tool call). Default true.
autoCapture	store salient user statements (preferences / decisions / facts) automatically. Default true.
captureMaxChars / captureMaxPerTurn	skip auto-capturing messages longer than this (default 2000) / cap memories stored per turn (default 3)
storageOptions	cloud credentials keyed by object_store config strings — S3 aws_access_key_id/aws_secret_access_key/aws_region (+ aws_endpoint for R2/MinIO), Azure azure_storage_account_name/azure_storage_account_key; omit for ambient cloud identity
recallK / recallMaxChars	results per recall (default 8) / max query chars embedded (default 1000)

Supported embedders today: local (the default — in-process Hugging Face transformers.js, all-MiniLM-L6-v2, 384-dim, no API key and no external service), Ollama (provider: "ollama"), and OpenAI-compatible (provider: "openai" + apiKey/baseUrl). Reusing OpenClaw's configured provider auth profiles is a future enhancement.

Local testing — no publish required

Layer 0 — the store, without OpenClaw (fastest signal):

npm install      # installs deps incl. @infino-ai/infino from public npm
npm test         # builds + runs deterministic oracle tests on real infino

Layer 1 — inside OpenClaw, via local link:

npm run build
openclaw plugins install --link "$(pwd)"   # references this dir; no copy
openclaw plugins enable memory-infino
# set plugins.slots.memory = "memory-infino" and an embedder in config (above)
openclaw                                     # restart the gateway; edits apply on restart

Scope

Hybrid recall + auto-recall / auto-capture + object storage, via the three standard memory tools. Deliberately out of scope for now: cross-encoder reranking, multi-scope (per-user/session) isolation, and a management CLI.

Layout

src/
├── infino-store.ts   # store on infino: hybrid recall (RRF) / SQL / forget / optimize / gc  (oracle-tested)
├── auto.ts           # auto-recall + auto-capture lifecycle hooks (+ pure helpers)
├── config.ts         # config schema + OpenClaw→infino mapping
├── embeddings.ts     # local (default, transformers.js) + Ollama + OpenAI-compatible embedders
└── index.ts          # OpenClaw plugin entry: the three tools + installs the auto hooks
test/
├── infino-store.test.mjs   # store oracle tests
└── auto.test.mjs           # auto-recall/capture helpers + end-to-end hook flow
openclaw.plugin.json  # plugin manifest (kind: memory, tools, config schema)