CodeAbra/iai-personal-memory-engine v1.0.0 on GitHub

First stable release. The architecture has settled and the public surface — the MCP tool set and the on-disk store — is committed-to from here on.

Highlights

Hippo storage engine — one encrypted local store holding records, the vector index, the graph, and the event ledger together, on SQLite + hnswlib + AES-256-GCM.
Native Rust engine — the text embedder and the graph kernels run as a compiled Rust extension, built automatically during install. No large Python ML runtime.
MOSAIC community detection — an original, MIT-licensed pure-Python + Numba algorithm written for the memory-graph workload (no GPL graph dependency).
Lilli HD substrate — hyperdimensional representations backing the episodic / semantic / procedural tiers, with structural recall at zero LLM cost.
Cross-session episodic recall — turns are captured verbatim and a relevant slice is surfaced at the start of each new session; also queryable through the iai CLI.
Subscription-billed consolidation — the nightly LLM step runs through your existing Claude subscription (claude -p); no API key, capped at ≤1% of the daily quota.
Export / backup / restore CLI, a write-ahead log for destructive sleep operations, a typed exception hierarchy, and iai-mcp doctor (23 health checks).

The previous embedded vector database, the PyTorch embedding stack, and the third-party hyperdimensional-computing and graph libraries — all out of the runtime path. No GPL dependency, no paid-API path.

git clone https://github.com/CodeAbra/iai-personal-memory-engine.git
cd iai-personal-memory-engine
pip install .

pip install builds the native Rust engine automatically via setuptools-rust.

Upgrading from a 0.x store — import once before the first 1.0.0 start:

pip install ".[migration]"
python scripts/migrate_lance_to_hippo.py

The migration script backs up the old data and verifies it byte-for-byte before removing anything.

Full details in CHANGELOG.md.