Skip to content

Releases: KanishkNoir/cognikernel

Release list

v0.1.0 — first public release

v0.1.0 — first public release Latest
Latest

Choose a tag to compare

View all tags
@KanishkNoir KanishkNoir released this 23 Jul 18:15
v0.1.0
354a41f

First public release. 🎉

CogniKernel gives Claude Code and Codex persistent, structured project memory — the decisions, constraints, and abandoned approaches worth keeping are captured automatically and injected back at the start of each session, so the agent stops re-deciding what you already decided. No LLM in the loop: memory is built by two small fine-tuned encoder models running locally on CPU in milliseconds.

Install

pip install cognikernel
cognikernel init .
cognikernel install-heads   # fetches the fine-tuned encoder bodies (~255 MB, sha256-verified)

On PyPI: https://pypi.org/project/cognikernel/0.1.0/

What's inside

  • Event-sourced SQLite memory (typed decisions / constraints / do-not-retry graveyard) with latest-wins supersession
  • Lexical-primary hybrid retrieval (FTS5 BM25 ∪ optional dense → RRF)
  • Four fail-open hook surfaces + an AST skeleton ranked by PageRank
  • Cross-platform: one store shared between Claude Code and Codex
  • Two fine-tuned encoders (salience + supersession) — classification, not generation, so zero per-session API cost

Schema v18 · CI: lint + full suite on Ubuntu (3.11/3.12) + Windows.

Assets 2
Loading

Encoder heads v1 (salience_v2 + supersession_xenc)

Encoder heads v1 (salience_v2 + supersession_xenc)

Choose a tag to compare

View all tags
@KanishkNoir KanishkNoir released this 18 Jul 08:15
heads-v1
d66feaa

Trained ONNX encoder artifacts consumed by memlora install-heads:

  • salience_v2-body.onnx / salience_v2-tokenizer.json — the fine-tuned SetFit encoder body for v2 salience extraction (the small classification head ships in-repo as src/memlora/extraction/heads/salience_v2.npz).
  • supersession_xenc-body.onnx / supersession_xenc-tokenizer.json / supersession_xenc-threshold.json — the supersession cross-encoder and its precision-calibrated decision threshold.

SHA-256:

551dff21e4bef79a7cc6f7f196d6f8a836a9752cce49617e20cdced93907f7a8  salience_v2-body.onnx
de2a1b8d1aee816739bc3ca0125a473a080bf05311c640f63bc22099527b1c6b  salience_v2-tokenizer.json
0eed82dbda0c3da3fcb64e1c6175602df106a92479ec0e06aef8f79553378ef6  supersession_xenc-body.onnx
851ca67100d372ca3ae031a6abd168f53489eebfd7d89523f35c5c9b4d372c3c  supersession_xenc-tokenizer.json
cc6087b344be54772989e8b6264800f7d2dc09d944080ee74fd5c58f0911dc9d  supersession_xenc-threshold.json

Both heads are optional: without them the system fails open to legacy salience / lexical supersession. Regenerable from scripts/train_salience_setfit.py, scripts/train_supersession_xenc.py and the export scripts.

Loading