built a memory layer for AI agents that is local-first, structured, and — unusually for this space — benchmarked through the real production retrieval pipeline, with a public runner so you can reproduce the number yourself.
The pitch in one line: give Claude Code / Cursor / Windsurf / any MCP agent persistent memory across sessions — no account, no cloud, no API key, data stays on your machine as Markdown.
Why it's different from the usual memory tools:
Workflow rules injection, not manual orchestration. When you run
npx @awareness.market/setup, it writes rules into your IDE config (CLAUDE.md, .cursor/rules/). The agent automatically callsawareness_initat session start,awareness_recallbefore work,awareness_recordafter changes. Mem0/Zep require you to manually wireadd()/search()at the right moments — if you forget, memory doesn't happen.Structured knowledge, not snippets. 13 categories: decisions, problem-solutions, workflows, pitfalls, skills, preferences, risks, action items. With conflict detection and dedup built in. A decision isn't a sentence floating in a vector index.
Benchmarked honestly. 96.0% R@5 on LongMemEval (500 questions) — measured through the real
unifiedCascadeSearchpipeline, not an independent harness. The runner is public and per-question isolated:
benchmarks/longmemeval/run_f053_daemon_path.mjs
For context: Mem0's open-source edition independently tested ~32-49% on the same benchmark (vendor self-reports 93.4% on managed), Zep 63.8%, MemPalace 96.6% (raw verbatim).
- Local-first is a feature, not a compromise. Zero LLM calls on retrieval, works fully offline, hybrid FTS5 + vector. Cloud sync optional when you want team collaboration.
Stack: Node.js daemon (SQLite + FTS5 + ONNX embeddings) · MCP native · 13+ IDE support · Python/TypeScript SDKs with interceptor pattern for OpenAI/Anthropic clients · ERC-8350 on-chain anchoring of memory state (yes, the memory system itself anchors its state to Sepolia).
Try it:
npx @awareness.market/setup
Full comparison: https://awareness.market/sdk-docs/ALTERNATIVES.md
Benchmark details: https://awareness.market/sdk-docs/LONGMEMEVAL.md
SDK repo: https://github.com/everest-an/Awareness-SDK
Happy to answer questions about the retrieval architecture, the benchmark methodology, or the anchoring design.
Top comments (0)