Skip to content

v0.43.0 — Claude Code CLI Automatic Memory Technical Preview

Pre-release
Pre-release

Choose a tag to compare

View all tags
@LiLara-AI LiLara-AI released this 29 Oct 12:20
· 11 commits to main since this release
v0.43.0
8bb48fc
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

ShadowGraph 0.43.0 — Technical Preview

This release adds automatic experience memory for the Claude Code CLI on Windows,
evaluated in print mode (claude -p). After a one-time setup, ordinary Claude Code
CLI sessions are captured durably, extracted in the background by a separate,
owner-selected Claude model, and delivered to fresh sessions at start and with
relevant prompts. This is a pre-release, not stable/GA. The package remains
private: true and unpublished to npm. Earlier tags and assets (v0.41.0 to v0.42.3)
are preserved. The Windows Codex workflow and the MCP Full, Compact and Workflow
profiles are unchanged.

Changes

  • Durable capture spool for Claude Code hooks: input is kept while the store is busy,
    replayed oldest first, and losses are recorded as content-free markers that
    delivery reports.
  • Runtime-evidence extraction profile for Claude Code 2.x: each invocation's own
    session account must show the confined session before its output is used; the
    owner chooses the extraction model at activation (--model).
  • Corrections on the Claude route: a user's correction supersedes the user's earlier
    decision; a decision kept verbatim as a memory is linked to the correction
    (applied: false) and marked correctedBy, not closed.
  • shadowgraph sharing project|selected|all-permitted: owner-confirmed, revocable
    delivery grants.
  • Delivery puts lines before full records at SessionStart and UserPromptSubmit, so one
    large record no longer crowds out the decision that says what was rejected and why.
  • An extraction output refused by the executor's own checks is asked once more,
    within the frozen budgets (4 calls per drain, 12 per rolling hour).
  • A decision organized from a whole message keeps its choice and alternatives when
    the proposed reason names none of them; the reason stays in the quoted source.
  • Captures from an earlier extraction setup are reported as priorSetup with a
    prior_setup gap, not as pending work. They are not extracted by a new activation.

The full list is in the changelog.

Install

npm install --global https://github.com/LiLara-AI/shadowgraph/releases/download/v0.43.0/shadowgraph-unified-plugin-0.43.0.tgz

Verify the tarball against SHA256SUMS.txt on the
GitHub Release.
Installation alone does not activate automatic memory. Claude Code setup, activation
and its commands are in the Claude Code guide.
Stop active work and back up stores and private integration state before upgrading.
Schema 7 is unchanged. See upgrade and rollback.

Evidence

On Windows 11 with Claude Code CLI 2.1.312 in print mode, a fresh-store run showed
one complete chain: a session failed a real check, repaired it, and kept the decision
with the rejected alternative; extraction ran automatically; a fresh session (Arabic
prompt) received that decision at SessionStart and used the same exact-arithmetic
approach before its first check, where a no-memory control used floating point. That
session still made and repaired a type error. Recovery of pending extraction by the
next ordinary session under the same activation was also shown. The Claude Code
guide lists each run.

Known limits

  • Delivered experience is not always applied. In one live run a fresh session
    received an earlier lesson, named it, and implemented only what its own sample
    showed. The cause is not established.
  • After extraction is deactivated and activated again, items left pending under the
    earlier setup are not extracted. Status reports them as priorSetup; this
    reporting was tested offline and in CI, not in a live run.
  • A user's decision may be kept as verbatim memory under strict claim verification.
    The separate reasonRejected field can be empty while the reason stays in the
    quoted alternative and source.
  • Extraction quality varies by model: Sonnet was reliable in the live runs; Haiku
    once missed the decision.
  • With Claude Code 2.1.302 or earlier on Windows, claude -p exits only after the
    extraction worker finishes (up to about 4 minutes). Use 2.1.312 or later.
  • Not verified: Claude Desktop, the interactive TUI, resume flows, and macOS or Linux
    hosts. No claim of universal model obedience, general bilingual accuracy, cost or
    time savings, or superiority over native memory.
  • Automatic extraction sends selected captured data to the configured model
    provider and uses its account. Pattern redaction cannot guarantee every secret is
    removed.
  • The credential-free Glama MCP container does not capture or extract a user's
    Claude Code or Codex sessions.

Assets

shadowgraph-unified-plugin-0.43.0.tgz and SHA256SUMS.txt. Built with npm pack from a clean export of 8bb48fc3cc65b9dd9e1a04cb6d9fc2117f90cc35.

Assets 4
Loading