Compare workflows for durable project rules, context layering, agent memory, isolation, and auditable handoffs.
The job
Includedproject instruction files and conventions · agent context layering and memory governance · cross-session handoffs and project isolation
Not this pagepersonal knowledge-base organization · secret storage · guaranteed autonomous long-term memory
30-second route
It directly reviews captured memory, distinguishes scratch context from project rules, and promotes approved recurring patterns into durable or scoped instructions.
Editorial confidence: highIt explicitly inventories governance surfaces, handles canonical-versus-alias precedence, and produces validation and coverage artifacts.
Editorial confidence: highIts explicit hierarchy, selective loading patterns, and session-management guidance address focused context without conflating it with permanent memory.
Editorial confidence: highThose needs are outside the task scope, and the supplied evidence does not establish such guarantees.
Editorial confidence: highBefore you install
Only Self Improving Agent’s supplied source documents installation commands. The other two provide source-backed workflows and outputs but no installation procedure in the supplied SKILL.md.
Read the intervention guide →Source-led comparison
Facts are extracted from each selected Skill’s declared workflow. Signals describe public adoption; they are not a universal score.
Source notes
Why these candidatesEach candidate comes from tracked public sources and has a source-attributed adoption signal. This is a shortlist, not a catalogue or global ranking.
What this cannot establishWe read declared instructions and public signals. We do not run skills, benchmark outputs, audit security, or claim a universal best choice.
Use it forThe strongest default for governing project memory across sessions: review captured learnings, promote durable ones into project rules, and keep scratch memory from becoming stale policy.
Avoid ifAvoid as the sole solution when you only need to author or audit repository instruction files and do not use a compatible memory system.
ActionInstall and use its memory-review and promotion lifecycle to graduate recurring, approved project learnings into durable rules.
EditsSeparate temporary observations in MEMORY.md from enforced project rules in CLAUDE.md or scoped .claude/rules/ files; remove promoted entries from scratch memory.
SetupClaude Code plugin installation is documented; the source also provides a Codex CLI installation command.
ContextIt explicitly distinguishes automatically captured project learnings, project-wide rules, scoped rules, and session continuity, with an approval step before promotion.
OutputMemory-health findings, promotion candidates, durable rules, and optionally extracted reusable skills.
ControlHuman-approved promotion: a pattern is flagged after recurrence and is promoted only after user approval.
Invocation/si:memory-review, /si:promote, /si:memory-status
Install scopeClaude Code plugin; Codex CLI script is also documented
Agent fitClaude Code has full stated support; OpenClaw and Codex CLI are described as adapted integrations.
Watch outThe workflow is centered on Claude Code auto-memory; its Codex CLI support is described as adapted rather than full support.
The skill curates auto-memory into durable project knowledge and promotes patterns into CLAUDE.md and .claude/rules/.SKILL.md:lines 8-10 · inspected 2026-07-20 · high confidence
Its documented commands review memory, promote patterns, extract skills, report memory health, and explicitly remember knowledge.SKILL.md:lines 12-20 · inspected 2026-07-20 · high confidence
The source defines a promotion lifecycle with recurrence, user approval, promotion to rules, and removal from MEMORY.md.SKILL.md:lines 73-80 · inspected 2026-07-20 · high confidence
The source explicitly contrasts MEMORY.md background context with CLAUDE.md enforced instructions and scoped rules.SKILL.md:lines 95-120 · inspected 2026-07-20 · high confidence
Platform support and installation paths, including Codex CLI, are documented.SKILL.md:lines 43-59, 149-156 · inspected 2026-07-20 · high confidence
Use it forRepository instruction-file governance: establish a canonical instruction surface, manage aliases, detect policy drift, and produce an auditable coverage map.
Avoid ifAvoid when the central need is automatically curating cross-session working memory rather than governing documentation and instruction files.
ActionAdapt its governance-document workflow to inventory AGENTS.md, CONTRIBUTING.md, aliases, and documentation surfaces; then reconcile precedence and drift.
EditsChoose a canonical instruction file when present, retain aliases as compatibility surfaces, and document precedence and conflicts.
SetupNo installation command is present in the supplied source; use the workflow only after confirming how the host makes the skill available.
ContextThe workflow treats agent instructions as part of a full documentation governance surface and explicitly includes canonical-versus-alias handling.
OutputUpdated documentation or review findings, validation notes, an instruction-surface map, coverage map, and remaining gaps.
ControlStructured build-or-review workflow with explicit inventory, reference reading, validation, and optional bounded delegation.
InvocationNot yet mapped from the source.
Install scopeNot yet mapped from the source.
Agent fitDesigned for agent instruction files and explicitly accounts for Codex, Claude, and Cursor handling in its output map.
Watch outIt is a documentation-governance workflow, not a dedicated persistent-memory capture or retrieval system.
The skill's purpose includes contributor-governance and agent instruction files.SKILL.md:lines 11-13 · inspected 2026-07-20 · high confidence
It covers AGENTS.md, CONTRIBUTING.md, aliases, canonical treatment, and agent-file drift.SKILL.md:lines 20-25 · inspected 2026-07-20 · high confidence
Its workflow inventories governance and documentation scope, reads rules for canonical/alias mapping and precedence, and returns validation notes and gaps.SKILL.md:lines 27-43 · inspected 2026-07-20 · high confidence
Its outputs include an agent instruction-surface map with a Codex/Claude/Cursor handling plan.SKILL.md:lines 68-79 · inspected 2026-07-20 · high confidence
Use it forA layered context method for projects that need focused session startup, task-specific context loading, safe handling of untrusted inputs, and deliberate session handoffs.
Avoid ifAvoid if you specifically need a packaged system to capture and promote learned memory across sessions; this is a method for structuring context rather than a memory store.
ActionAdapt the hierarchy as the project’s context policy: keep durable rules always loaded, load specs and source selectively, pass only relevant errors, and summarize before handoff or compaction.
EditsCreate a concise rules file, maintain a project-map index for large codebases, load only the applicable spec and source files, and refresh sessions when major work changes.
SetupNo installation command is present in the supplied source; adopt the documented hierarchy and checklists in the project’s existing instruction system.
ContextIt defines persistence-to-transience layers and concrete techniques for selective inclusion, hierarchical summaries, context refresh, and ambiguity handling.
OutputA rules-file baseline, selectively packed task context, project-map summaries, and verification checks.
ControlLayered and selective: persistent rules first, then per-session/per-task materials; surface conflicts instead of silently choosing.
InvocationNot yet mapped from the source.
Install scopeNot yet mapped from the source.
Agent fitNames CLAUDE.md and equivalent rules-file locations for Cursor, Windsurf, GitHub Copilot, and OpenAI Codex.
Watch outThe hierarchy is guidance; it does not itself promise autonomous long-term retention.
The source defines context engineering as deliberate curation of what agents see, when they see it, and how it is structured.SKILL.md:lines 8-10 · inspected 2026-07-20 · high confidence
It provides a hierarchy from always-loaded project rules through session, task, and iteration context.SKILL.md:lines 20-40 · inspected 2026-07-20 · high confidence
It names equivalent instruction-file locations for Cursor, Windsurf, GitHub Copilot, and OpenAI Codex.SKILL.md:lines 74-79 · inspected 2026-07-20 · high confidence
It directs users to start fresh sessions, summarize progress, and compact deliberately as context grows stale or long.SKILL.md:lines 113-120 · inspected 2026-07-20 · high confidence
It verifies rules-file coverage and context refresh when switching major tasks.SKILL.md:lines 282-290 · inspected 2026-07-20 · high confidence
Still comparing?