V5.1.1: Code memory joins documents and history

V5.1.1: code memory joins documents and history
DaiDocs now carries a code tier alongside its document and session memory, so one system answers across code, docs and history. The version unifies on 5.1.1 (up from 4.4.36) to mark that line. The benchmark provenance (PROVENANCE V4.4n32), FORMAT (4.4) and the engine id are unchanged, so every cited benchmark number stays valid.

What is new
Kerneta-Cai code tier
A plain-text .cai code graph (10 languages, tree-sitter) with document cross-links and a session-history reader, sharing one ask router with the .dai memory.

pip install kerneta-cai (Python package).
It stays off the npm whitelist, so npm install daidocs is unchanged.
node setup.js --ask offers to install it in the same flow.
kerneta setup --global --auto makes an agent query it automatically in every project.
/recall and /remember slash commands for Claude Code
Two prompt templates over the daidocs-mcp tools:

/recall searches your .dai memory and answers following the reading protocol.
/remember saves a note, writing the extraction itself so it needs no API key. Installed by node setup.js into ~/.claude/commands/ (or project-scoped), and tracked like every other surface, so --status reports them and --restore removes them.
Small-store coverage floor
For stores up to about 20k tokens the reader keeps the selective deep read and also includes every session's compact summary, so an answer no longer depends on the ranker picking the exact session. Large-store behaviour and the published LongMemEval numbers are unchanged.

Reproducible code benchmarks
code-benchmarks/ is a self-contained, offline kit that rebuilds the .cai vs Graphify vs raw-files numbers from source with python code-benchmarks/run_all.py (no API key, no network). On psf/requests it reproduces .cai 16/16 at about 55 tokens a query against raw's 9,023 (164x fewer) and Graphify's 689 (13x fewer); on httpx structural queries .cai holds 100% recall at a fraction of the tokens (imports 218x fewer than Graphify, subclasses 73x). Kerneta-Cai/BENCHMARK-REPRODUCE.md documents how to rebuild a store correctly.

Install or upgrade
npm install -g daidocs # new install
daidocs update # upgrade an existing install, then restart the host app
pip install kerneta-cai # optional code tier
Full details are in CHANGELOG.md, the [5.1.1] section

V4.4n32: launch release

The launch release. No engine change: the method is byte-identical to the build the release run measured, and every benchmark artifact hashes to the same values. Every number below the fold is the same number. What changed is the product around it.

After npx daidocs setup, you are good to go: if you bring an API key it is used straight away (all local on your own system), and on a Claude subscription no key is needed at all. Nothing else to configure.

Added

• Memory that lives with the work. A folder can be declared a project and keep its own store that travels with the code. Seven folder types decide what happens to that memory: normal, locked, frozen, connected, shared, confidential, temporary.

• Sessions that save themselves. The Stop hook converts every 4,000 new tokens, written by the assistant already in the conversation. Below the threshold the session is held verbatim and resumes next time the folder is opened.

• Subscription mode. On a Claude host no background path calls a paid model: where an extraction is needed the assistant is asked for it, not billed for it.