Reference implementation for the iHow Memory spec.
Apache-2.0 · Local-first · File-protocol · Passes spec v0.1 conformance
The reference implementation behind the iHow Memory spec. Use it as:
- A working memory layer for your AI agent (file-protocol, no SaaS, no API keys)
- A baseline to compare your own memory system against
- The implementation behind the public LongMemEval_S 470/470 retrieval-stage benchmark
To try it from a clone today:
git clone https://github.com/iHow1/ihow-memory-core.git
cd ihow-memory-core
npm install
npm exec --package . -- ihow-memory init my-project
cd my-project
# Read README.md to see the generated workspace layoutA polished published-package quickstart is on the v0.2 roadmap. Meanwhile the repo-local invocation works.
- Local file workspace as source of truth
- Append-only audit log (
memory/_events/*.ndjson) - Scoped memory per agent runtime (Claude Code / Codex / OpenClaw / your own)
- Retrieval via local vector + lexical hybrid with cluster-aware rerank
For full spec, see iHow Memory standard.
| Benchmark | Result |
|---|---|
| LongMemEval_S (470 effective samples) | recall_all@10 = 1.0 |
| 5 spec scenarios v0.1 (self-conformance) | 5/5 PASS |
Important: this is retrieval recall, not end-to-end answer accuracy. For per-layer comparison context and methodology, see the spec repo and its evidence manifest.
Some techniques used in our benchmark harness port directly to production retrieval (original-query anchor preservation, cluster-aware rerank, bounded sibling expansion). Others are LongMemEval-shaped heuristics that remain benchmark-only.
The public evidence manifest summarizes this boundary. Benchmark-only heuristics are not packaged as general retrieval improvements.
Apache-2.0. Spec (CC-BY) lives in the standard repo.
- Bug reports: issues
- Discussion: GitHub Discussions
- Spec proposals: standard repo