Skip to content

Releases: ozgurkarahan/ai-agent-memory

v2026.07.1 — Operationally self-contained bootstrap

Choose a tag to compare

@ozgurkarahan ozgurkarahan released this 24 Jul 10:22

Makes every workflow installed by bootstrap.md executable using only files the bootstrap creates—no private helper scripts or maintainer-specific paths.

Highlights

  • Zero-helper-script runtimeingest, lint, and review-sessions now use agent-native file and JSONL operations.
  • Portable installation — all nine workflows resolve the wiki root from the installed structure or AGENT.md; new-engagement resolves its destination and template instead of assuming a home-directory layout.
  • Offline project scaffolding — the complete 9-file project-template/ is now a required bootstrap deliverable.
  • Fail-fast dependency checks — generation rejects tri-surface drift, unshipped scripts/*.py references, and maintainer-specific paths.
  • Clean-room proven — independent agents installed all 27 skill files and completed an end-to-end ingest + lint; an external hash baseline confirmed the activity event count matched every changed path.

Existing installations can rerun the new bootstrap.md safely: matching files are preserved and missing or outdated workflow files are repaired.

See CHANGELOG.md for the full list.

Full diff: v2026.07.0...v2026.07.1

v2026.07.0 — Reliable skills-first bootstrap

Choose a tag to compare

@ozgurkarahan ozgurkarahan released this 17 Jul 12:54

Makes the self-contained bootstrap reliable for no-code, agent-only installation and clarifies how every coding agent connects to the shared memory loop.

Highlights

  • Skills install first — the bootstrap creates the 9 Copilot CLI instructions, 9 Claude Code skills, and 9 agent-neutral workflow copies before building wiki content.
  • Fail-closed Definition of Done — interrupted runs resume safely, and a wiki-only partial installation cannot report success. Completion requires all 27 skill files with matching bodies.
  • Clean-room verified — two independent coding agents completed the bootstrap using only bootstrap.md; both produced the full wiki, skills, root instructions, and project template.
  • Karpathy memory-loop diagramREADME.md now visualizes how knowledge compounds from one coding session into the next.
  • Shared wiki pointersAGENTS.md and .github/copilot-instructions.md direct agents to consult the central wiki before answering from memory.

See CHANGELOG.md for the full list.

Full diff: v2026.05.2...v2026.07.0

v2026.05.2 — Complete project-template + resync bootstrap.md

Choose a tag to compare

@ozgurkarahan ozgurkarahan released this 13 May 16:39

Completes the public-facing setup. A fresh clone now ships a viable project-template/ for new-engagement, and bootstrap.md installs the full 9 skills × 3 surfaces that the rest of the repo already provides.

Highlights

  • project-template/ completed — 5 missing canonical files added (LICENSE, README.md, .claude/CLAUDE.md, .claude/commands/status.md, .github/instructions/end-session.instructions.md) so new-engagement works out of the box on a fresh clone.
  • bootstrap.md fully re-synced (~80 KB, 1939 lines) — now installs 9 skills × 3 surfaces (was 3 skills × 1 surface). Each of ingest, end-session, query, lint, plan-week, close-week, project-status, review-sessions, new-engagement is emitted as .github/instructions/{slug}.instructions.md + .claude/skills/{slug}/SKILL.md + memory/workflows/{slug}.md. Also adds the full 9-file project-template/ scaffold as optional Step 11 + memory/ops/ infrastructure for plan-week/close-week.
  • scripts/regenerate-bootstrap.py — generator that assembles bootstrap.md from the canonical skill bodies and project-template/ files. Run it after any skill or template change to keep bootstrap.md in sync.
  • YAML hotfix for all 9 .claude/skills/*/SKILL.mddescription: fields are now properly double-quoted so GitHub's frontmatter parser stops failing with mapping values are not allowed in this context.
  • new-engagement skill (all 3 surfaces) — replaced the maintainer's private wiki reference with a concrete Prerequisites block tailored for public users.

See CHANGELOG.md for the full list.

Full diff: v2026.05.1...v2026.05.2

v2026.05.1

Choose a tag to compare

@ozgurkarahan ozgurkarahan released this 13 May 16:06

[2026.05.1] — 2026-05-13

Adds Claude Code as a first-class agent surface and polishes the surrounding docs.

Added

  • .claude/skills/{slug}/SKILL.md surface — Claude Code agents working inside a clone of this repo now auto-discover all 9 skills via Claude Code's project-scoped skill discovery (.claude/skills/). Each SKILL.md carries the YAML frontmatter Claude Code expects (name, description) so the skill router can match user requests against the right skill. The repo is now tri-surface and treats GitHub Copilot CLI and Claude Code symmetrically:
    • .github/instructions/{slug}.instructions.md — Copilot CLI
    • .claude/skills/{slug}/SKILL.md — Claude Code
    • memory/workflows/{slug}.md — agent-agnostic plain-Markdown reference
  • .gitignore exception pattern (.claude/* + !.claude/skills/ + !.claude/skills/**) so the public skills surface ships with the repo while personal Claude Code state stays ignored.

Changed

  • CONTRIBUTING.md — "dual-surface" guidance updated to "tri-surface", with a table summarising each surface's frontmatter requirements and rationale.
  • README.md — "How It Works" rewritten (was "Three workflows drive the system", now reflects 9 skills); "Agent Compatibility" tables add rows for GitHub Copilot CLI (.github/instructions/) and Claude Code project-scoped skills (.claude/skills/).

Fixed

  • plan-week / close-week skills (both .github/instructions/ and memory/workflows/ surfaces) — removed leaked references to ~/.claude/skills/{slug}/SKILL.md (mirror) — abridged here:. With the new .claude/skills/ surface now shipping in the repo, each surface is self-contained — no dangling pointer needed, and the procedure flows directly from the trigger sentence.

Notes

v2026.05.0 — 9 agent-agnostic skills + schema refresh

Choose a tag to compare

@ozgurkarahan ozgurkarahan released this 13 May 15:42
ad9cd53

First versioned release.
Full history: CHANGELOG.md
First versioned release. Brings the new and updated agent skills from the maintainer's private memory wiki back to this public teaching repo, scrubbed of real client / colleague / personal references.

Added

  • 9 agent-agnostic skills, each shipped in two parallel locations to satisfy both conventions used in this repo:

    • .github/instructions/{slug}.instructions.md — for GitHub Copilot CLI (applyTo: "**" frontmatter)
    • memory/workflows/{slug}.md — canonical plain-markdown procedure (the existing repo convention)

    New skills:

    • close-week / plan-week — Weekly operating rhythm using ops/weekly/{ISO}.md files + ops/activity.jsonl event stream
    • project-status — 30-second briefing on any child project (tech stack, history, deployment, current state)
    • review-sessions — Session-data analysis for workflow efficiency, prompt quality, error patterns. Agent-agnostic: supports GitHub Copilot CLI (~/.copilot/session-state/{id}/events.jsonl) and Claude Code (~/.claude/projects/{hash}/{id}.jsonl) via --agent {copilot,claude,all}
    • new-engagement — Scaffold one or more client engagement projects from ~/projects/project-template/
    • lint — Wiki health check trigger
  • Schema updates in memory/schema.md:

    • 11 new category rows to the taxonomy table: Clients, Skills, Agents, Tools, Meetings, Career, Microsoft, Personal, Queries, Ops, Agent Config
    • Client-folder rule for wiki/projects/ with a slug-strip decision table (when to flatten client-foo slugs vs. keep them under client/foo/ sub-folders)
    • Source classification frontmatter fields: source_type, sensitivity, system_of_record, retention, sharing

Changed

  • ingest skill — major refresh: adds Phase 0 (inventory), Phase 2b.5 (consolidation matrix for mass ingest), Phase 6 (7-item self-audit gate), Phase 8 (emit activity event for cross-skill operational tracking)
  • end-session skill — aligned with the post-.ai/ world: project knowledge now lives directly on the wiki page rather than in per-project .ai/ folders
  • query skill — tightened workflow

Removed

  • launch-copilot / launch-claude skills — public consumers shouldn't have to download and run wrapper scripts from an internet repo, and the wrappers added no value over the equivalent native CLI invocation (cd <project> + copilot --allow-all-tools / claude --dangerously-skip-permissions). The new-engagement skill was updated to use native commands in its examples.
  • sync skill — targeted a .ai/ folder convention that was retired in April 2026 (project knowledge now lives directly on the wiki page, see end-session); the reference script was not shipped publicly and contained internal client-specific hardcoded mappings. If a generic drift-detection workflow is wanted later, it should ship as a fresh skill paired with a configurable script.

Notes

  • All examples were scrubbed via a one-shot script + manual review: client names → Acme / Contoso / Fabrikam / Northwind / Globex, colleague names → Colleague A..I, personal paths → ~/projects/..., repo refs → github.com/{your-username}/..., internal back-references → internal-*-lesson placeholders.
  • review-sessions defines an agent-agnostic contract but no reference parser ships in this release — implementers should match the schema documented in the skill (Copilot CLI event-stream parser + Claude Code turn parser).
  • Reference: PR #1.