KAI v7.3.2

KAI v7.3.2 — memory-safety guards (patch)

Triggered by a knowledge-loss incident: a CWD mismatch plus an unguarded memory-file deletion lost days of accumulated work. This patch is fixes + recurrence guards only — net-new capabilities are deferred to 7.4.0.

Guards

• Memory-file deletion protection — SecurityValidator now blocks rm and mv of memory files (anchored to .claude/(projects/*/memory|MEMORY)/). The prior block only caught whole-directory rm.
• CRITICAL steering rule — never delete project memory based on format judgment.
• CWD-mismatch detection at SessionStart — warns when launched from a directory with no project marker, before work writes to the wrong tree.
• Sync atomicity — sync-to-kai writes a .sync-scrub-in-progress sentinel; pre-commit/pre-push refuse to run mid-scrub.

Docs

• CWD requirement documented in the memory system guide.

Full test suite green (1809 passing).

KAI v7.3.1

KAI 7.3.1 — Memcarry store-resolution hardening + release-pipeline fixes

A patch release tightening memcarry's store resolution and the release pipeline.

Fixes

• Memcarry canonical store-root resolver. The CLI and MCP server fell back to a package-relative seed directory whenever MEMCARRY_STORE was unset in a spawned process, silently diverging from the live store the hooks already used. A single resolveStoreRoot() in @memcarry/lib is now the source of truth (MEMCARRY_STORE → PAI_DIR/MEMORY/memcarry/store → HOME/.claude/...); CLI and MCP both resolve through it. Verified the resolver never resolves to the repo path under any environment combination.
• Release pipeline hardening. PII scrubbing now covers all-caps name variants (the detector was case-insensitive while the scrubber was case-sensitive), and the release verifier no longer trips on design-record references in specs/.

Surface: 69 skills · 57 hooks · 20 agents · Algorithm v3.14.0.

KAI v7.3.0

KAI 7.3.0 — Closing the cross-project knowledge cycle (Memcarry)

Memcarry is KAI's portable memory layer — it carries lessons learned in one project into the next. 7.2.0 shipped recall; 7.3.0 closes the loop with the two missing halves: backflow (refine global lessons from recurring project signals) and capture (turn a session's lesson into a reusable atom). The cycle is now complete: capture → store → recall → refine.

Backward-compatible feature release — nothing removed.

Surface: 68 skills · 57 hooks · 20 agents · Algorithm v3.14.0.

Keystone — the knowledge cycle

• Memcarry backflow (004). memcarry refine promotes recurring project signals into refined global lessons, with mid-turn refine plus an End-skill safety net wiring it into the session lifecycle.
• Memcarry capture (005). memcarry capture-lesson turns a session lesson into a stored atom; a steering rule and an End-of-session step surface it automatically.
• Hybrid recall hardening. Every-turn hybrid keyword+semantic recall into context, RRF fusion in recall(), a host-side score provider with a rebuildable vector cache, and re-injection of lessons after context compaction.

Statusline & telemetry

• Action tab — self-healing action line with git state and cached indicators.
• SkillTracker — skill-invocation telemetry for data-driven capability decisions.

Fixes & robustness

• Checkpoint sequence tiebreak for same-millisecond saves (CI flake).
• Semantic-fallback test stabilized against a Bun teardown crash.

Note: KAI ships the memcarry mechanism (lib + CLI + hooks). Lesson atoms are personal data — build your own store.

KAI v7.2.0

Consolidation + public-fork hardening.

• Embeddings/similarity engine unified (lib/similarity.ts, lib/embeddings.ts) — fixed a latent system-wide pooling bug that had silently broken semantic dedup and recall.
• Duplicate-flow elimination: merged ReadTracker+MemoryAccessTracker → ReadActivity; SessionEndComposite consolidates 10 SessionEnd hooks; shared transcript-cache and ratings-store.
• Robustness: StopOrchestrator graceful-shutdown deadline; board bound to loopback + mutation token; hook-wiring drift guard (reconcile-wiring.ts).
• Public-fork (KAI) hardening: single-source PII patterns + scrub; .tsx/.css now brand/PII-scrubbed (closed a live brand leak); portable sed for Linux; CI test-retry + ripgrep; checkpoint same-ms tiebreak.

Surface: 68 skills · 55 hooks · 20 agents · Algorithm v3.14.0. See CHANGELOG.md for full details.

KAI v7.1.0

Agent View integration + cognitive-persistence fixes.

• Two-layer architecture composing KAI's Algorithm inside Claude Code's Agent View.
• DevTeam background mode and /goal convergence for background sessions.
• Board AgentViewPanel UI + sessions API + SSE watcher; worktree isolation for key agents.
• Cognitive-persistence fixes: cross-project knowledge loss (hash normalization, index regeneration, content injection, instinct purge).
• INVESTIGATE mode added — prevents Algorithm ceremony on review/research/validation tasks.

See CHANGELOG.md for full details.

KAI v7.0.0

Meta-cognitive monitor and PlanningObserver integrated into Algorithm loop. CostTracker, StallDetector, adaptive retry, CheckpointManager wired into DevTeam orchestrator. PII patterns externalized. 1665 tests passing. See CHANGELOG.md for full details.

KAI 5.3.0 — Initial Public Release

KAI 5.3.0 — Personal AI Infrastructure for Claude Code

Built on Daniel Miessler's PAI system, extended with production hardening, multi-model inference, and individual deployment.

Highlights

• Algorithm v3.13.0 — effort tiers, ISC quality gates, confidence tags, capability selection
• 80 skills — Research, Security, Analysis, Writing, EM/PLM workflows
• 41 hooks — lifecycle automation with stderr wrapper + async flags
• 18 named agents — Architect, Engineer, 5 researchers, Pentester, and more
• PlanApprovalGuard — plan approval workflow with structured detection
• Deliberate research mode — parallel multi-model queries
• Interactive installer — 6-step setup wizard with domain-based config
• Security hardening — SecretScanner, GitHubWriteGuard, patterns.yaml

Install

git clone https://github.com/kai-cli/kai ~/kai
bash ~/kai/install.sh

Requirements

• macOS or Linux
• Claude Code CLI
• Anthropic API key (or AWS Bedrock credentials)

See QUICKSTART.md for the full getting-started guide.