Eggshell 0.1.0

Eggshell is local memory for AI agents: reuse original work and supporting evidence across independent chats, without LLM calls to organize memory.

This build preserves admitted observations while filling missing context in a selected explanation, avoids repeated graph-key preparation during packet fitting, and keeps provenance and complete packet bounds checked in Lean. Saved observations remain immutable; recording and indexing continue between chats.

LLVM measurement

Native macOS Codex CLI, GPT-6 Luna xhigh, independent ephemeral chats: empty memory → a new seed investigation → topics 1–3 with continuous saving → topics 1–3 with accumulated memory.

Topic	Fresh reference	First lap	Reduction	With accumulated memory	Reduction
1	1,284,909	224,815	82.5%	52,029	96.0%
2	799,753	170,441	78.7%	52,007	93.5%
3	2,302,471	657,499	71.4%	52,189	97.7%
Total	4,387,133	1,052,755	76.0%	156,225	96.4%

Input + output tokens, including cached input and reasoning output. The seed cost 766,235 tokens separately. All seven new memory-side chats completed; no failed attempt was excluded. One recorded independent fresh reference per topic is reused in both columns, so this is not a randomized or statistical savings guarantee.

The retained lap required no additional tool executions. The slowest measured hook was 23.6 seconds within its 30-second deadline. Every original observation and answer remained saved, with indexing queues empty before the next chat. Source review accepted the core answers, with minor fixture/citation qualifications recorded alongside the unedited answers; no passing LLVM runtime test or statistical quality equivalence is claimed.

Full measurement, handoffs, answers, and review notes · README

Version remains 0.1.0. Binaries and plugin metadata identify the same source build in runtime.json. Local full-suite, lifecycle and packaging checks, and macOS/Linux engine and adapter CI passed before merge; the release workflow separately builds and tests all four supported distribution targets.