amem

Monorepo for the amem agentic-memory stack — memories that evolve, not just accumulate. Qdrant + local Transformers.js + LLM, no Python required.

Packages

Package	What it is	npm
@amemhq/core	Framework-agnostic A-MEM engine — note construction, evolution, hybrid (BM25 + dense) retrieval with graph expansion. Qdrant + Transformers.js.	@amemhq/core
openclaw-amem	OpenClaw memory-slot plugin — a thin wrapper around amem-core.	openclaw-amem
amem-api	Thin single-writer service (HTTP + MCP) so multiple processes share one memory store.	coming soon

📖 Documentation: amem.owo.lc · 📄 Paper: A-MEM (arXiv:2502.12110, NeurIPS 2025)

Models

Two tiers, because the calls are not equally hard. fast runs everything frequent: extraction, link judgement, the per-turn CRUD decision. strong runs only merge adjudication and contradiction classification.

tier	env	plugin config
fast	AMEM_LLM_MODEL	llmModel
strong	AMEM_LLM_STRONG_MODEL	llmStrongModel

strong is optional and falls back to fast field by field, so setting only llmStrongModel keeps the same provider and endpoint with a better model, and setting all three llmStrong* fields runs the tiers on separate backends — a local Ollama for fast, a hosted API for strong. Set none and it behaves as a single-model install. There is no built-in strong default: an upgrade never starts spending more on its own.

The split is worth the config because the gap is uneven. Extraction differs about 2 points between a cheap model and a strong one; contradiction detection differs 17–21, and implicit contradictions collapse from 55% to 8.7%. Sources and the rest of the reasoning: Design Rationale.

Develop

This is a pnpm workspace (Node 24).

pnpm install                 # first run: `pnpm approve-builds` to allow onnxruntime-node / sharp / esbuild
pnpm -r build                # build every package
pnpm -r typecheck
pnpm -r test                 # vitest — integration tests need Qdrant on :6333 + ANTHROPIC_API_KEY
pnpm docs:dev                # run the docs site locally

Publishing is automated via Changesets + GitHub Actions.

License

MIT © heichaowo