Neonmem 0.9.7 is out.

#ai #crystal #programming #python

1. A two-level importer — two kinds of "stuff," treated differently

The big change. Your project doesn't come in one shape, so the importer no longer flattens
it into one pile:

Folders & files → a searchable knowledge pool. Your docs, code and notes are vectorised into a lossless, deduplicated facts pool — the same fact stated three ways becomes one fact, every source kept. Nothing is summarised away.
Agent chats → typed memories. Point Neonmem at a Claude (or other agent) transcript and it pulls out only what's worth keeping — the decisions, dead-ends and rules — as clean, typed memories. A decision is stored as a decision; a dead-end stays a warning. The process-narration ("I read the file…", "please check…") is dropped.
Links become knowledge. If a chat references a file on disk, that file is pulled into the pool automatically, with a memory that points back to it.

The result is labelled honestly in the UI: Facts loaded (the pool) and
Memories created (the kept decisions).

2. Grounded, offline recall

0.9.7 replaces the old embedder with IBM Granite-30M, run as a fused fp16 ONNX graph
through ONNX Runtime:

Database-class retrieval quality on any CPU — no GPU, no PyTorch, no API key, no cloud.
Every prompt walks memory in order — reflexes → short-term → long-term → facts pool — and answers from what you actually imported, or honestly says it doesn't know.

This is the headline behaviour: ask "what is ARC?" and you get your definition from
your docs — not the textbook expansion the model would otherwise guess. A memory that's
occasionally wrong is worse than no memory at all, so the rule is: answer from the user's
sources, or abstain. Never invent.

3. Tags that stick

Tag an import with a topic (e.g. Specific API) and Neonmem mints one clean, canonical
memory for it, linked back to the source — even when your docs never write the term
verbatim, as long as they clearly describe it. If the corpus genuinely has nothing on a
tag, it's left out rather than faked.

4. Clean by construction

Memories follow one golden rule: a single concise statement (ARC — your provisioning platform) linked to the full source, not a messy pile of raw chunks. Chat capture
deduplicates through the same facts layer, so re-importing a conversation never doubles up.

5. One durable cartridge

The importer keeps the full source corpus inside the cartridge (content-addressed + compressed) — one file replaces the scattered docs and transcripts, and the facts are always rebuildable from ground truth.
Opt-in AES-256-GCM encryption at rest — your whole corpus as a private vault.
Imported knowledge is long-term and survives reopening the project.

Built on (all open, permissively licensed)

Embeddings: IBM Granite-30M (Apache-2.0) via ONNX Runtime (MIT). Vector search:
FAISS (MIT). Agent integration: the Model Context Protocol. Full attributions ship
with every download. No third-party LLM, nothing phones home.