Skip to content
Skills Library
AI & agents · Planning

context-infrastructure

/context-infrastructure

Design, audit, or improve durable context and memory systems for AI agents using layered observations, reflection, conservative promotion, progressive disclosure, and replayable boundary artifacts. Use when building long-running agent memory, deciding what belongs in sessions versus rules or skills, preventing context bloat, or turning repeated corrections and workflows into durable assets. Do not use it to copy another person's private axioms or automatically promote one-off observations.

Repository
ax-skills ↗
Licence
MIT
Last repository change
16.10.2026

Assessment and security

90/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content25 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
2
Scripts
0
HTML comments
0
URLs in SKILL.md
2
Deprecation

No signals detected. Repository changed on 16.10.2026, without references to retired models or APIs.

Licence

MIT, open: use and modify while retaining author attribution.

What it does and when to use it

Design, audit, or improve durable context and memory systems for AI agents using layered observations, reflection, conservative promotion, progressive disclosure, and replayable boundary artifacts. Use when building long-running agent memory, deciding what belongs in sessions versus rules or skills, preventing context bloat, or turning repeated corrections and workflows into durable assets. Do not use it to copy another person's private axioms or automatically promote one-off observations.

Original author description

Design, audit, or improve durable context and memory systems for AI agents using layered observations, reflection, conservative promotion, progressive disclosure, and replayable boundary artifacts. Use when building long-running agent memory, deciding what belongs in sessions versus rules or skills, preventing context bloat, or turning repeated corrections and workflows into durable assets. Do not use it to copy another person's private axioms or automatically promote one-off observations.

How it works

  1. Define four layers
  2. Separate observation from promotion
  3. Route promotion by scope
  4. Design for progressive disclosure
  5. Make boundaries inspectable
  6. Protect integrity
  7. Completion contract
Original SKILL.md headings · 114 lines, 691 words
SKILL.md preview
---
name: context-infrastructure
description: Design, audit, or improve durable context and memory systems for AI agents using layered observations, reflection, conservative promotion, progressive disclosure, and replayable boundary artifacts. Use when building long-running agent memory, deciding what belongs in sessions versus rules or skills, preventing context bloat, or turning repeated corrections and workflows into durable assets. Do not use it to copy another person's private axioms or automatically promote one-off observations.
---

# Context Infrastructure

Treat context as a compounding system, not a giant prompt. Route, filter, test,
and promote experience so future agents receive the smallest useful context.

This AX adaptation is inspired by the public
[`grapeot/context-infrastructure`](https://github.com/grapeot/context-infrastructure)
reference implementation and the Observer/Reflector pattern described by
[Mastra Observational Memory](https://mastra.ai/research/observational-memory).
Reuse the architecture, not another person's accumulated axioms, preferences,
or private data.

## Define four layers

| Layer | Purpose | Typical contents |
|---|---|---|
| L0 Session | Execute the current task | Prompt, active diff, command output, current plan |
| L1 Observations | Preserve selected raw signals | User corrections, repeated friction, notable failures, explicit preferences |
| L2 Reflection | Distill and discard | Deduplicated patterns, contradictions, counterexamples, promotion proposals |
| L3 Durable assets | Change future behavior | Axioms, reusable skills, global methodology, technical knowledge, project memory |

Do not skip layers. Keep a new signal in L1 until reflection shows that it is
repeated, stable, useful outside the original task, and routed to the right
scope.

## Separate observation from promotion

Make the Observer append concise evidence. Make the Reflector merge, discard, or
propose promotion.

- Record only explicit corrections, preferences, retractions, and recurring
  workflow lessons worth reconsidering. Most tasks should produce no memory.
- Preserve source, date, scope, and counterexamples needed to evaluate a signal.
- Merge duplicates and flag contradictions during reflection.
- Generate a reviewable promotion proposal before changing durable rules,
  axioms, skills, or shared memory.
- Keep automatic promotion off by default unless the domain has deterministic
  validation, rollback, and an explicitly authorized policy.

This split prevents one intense session from rewriting long-term behavior.

## Route promotion by scope

| Repeated signal | Durable target |
|---|---|
| Stable personal preference | Personal axiom or preference profile |
| Repeated cross-project workflow | Reusable skill |
| Global operating methodology | Focused rule or policy |
| Reusable technical lesson | Knowledge entry |
| Project-specific decision or constraint | Project memory or source-of-truth documentation |
| One-off task fact | Keep ephemeral or discard |

Promote only claims that change future decisions. Preserve dissent and edge
conditions; a rule without boundaries is usually too broad.

## Design for progressive disclosure

- Keep the global entrypoint as a routing table, not the whole memory system.
- Use indexes to tell agents when to load focused rules, skills, or knowledge.
- Put triggering information in skill metadata and detailed procedure in the
  skill body or directly linked references.
- Retrieve project-specific context before global history when the task depends
  on current source truth.
- Prefer sparse, high-density context over broad dumps and repeated summaries.

Do not add another agent when a fixed-contract parser, index, or API call can
route the context deterministically.

## Make boundaries inspectable

For every context-producing workflow, define:

1. **Input boundary:** source records, time window, permissions, and freshness.
2. **Selection boundary:** inclusion, exclusion, deduplication, and privacy rules.
3. **Artifact boundary:** an inspectable observation or reflection record before
   AI judgment becomes durable.
4. **Promotion boundary:** owner, threshold, review, target, and rollback.
5. **Consumption boundary:** which task shape loads which durable asset.

Store AI-facing inputs and AI-produced intermediate outputs when they affect
visible or durable behavior. Debug bad output by inspecting freshness,
timestamps, filters, ordering, and fallbacks before changing the model or prompt.

## Protect integrity

Test the parts that future agents rely on:

- parser and schema shape;
- time-window and freshness filtering;
- deduplication and contradiction handling;
- fallback and retry signals;
- privacy, secret, and retention scans;
- deterministic replay from saved boundary artifacts;
- promotion thresholds and rollback behavior.

Minimize collection. Do not inspect private stores, upload records, or broaden
retention without authorization. Redact secrets before an observation becomes a
shared or durable artifact.

## Completion contract

Return a compact system design or audit containing:

- the L0–L3 stores and owners;
- Observer and Reflector contracts;
- the signal-to-target promotion table and review threshold;
- progressive-disclosure routing;
- privacy, freshness, replay, and rollback checks;
- remaining manual decisions.

Complete only when a new observation can be traced from source to consumption
without silent promotion or an uninspectable context jump.
Files in this package (2)
  • SKILL.md 5 KB
  • agents/openai.yaml 1 KB

Versions

The database has no other skills named context-infrastructure.

Database snapshot: 13.11.2026 · package version 41910469be47 (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 100company-agent-infrastructure
    Use when designing or building internal/company AI agents that need shared context across company systems, durable organizational memory across sessions or agents, user-scoped permissions, approvals, audit, or governed actions. Also use when deciding whether ordinary MCP tools,…
  • 100unhobble-agent-context
    Audits and designs lean, coherent context across system prompts, persistent instructions, repository guidance, skills, tool descriptions, memory, and references. Use for context engineering; creating, reviewing, simplifying, or reorganizing agent context; resolving conflicting,…
  • 95skill-architect
    Design, create, audit, and improve Claude Agent Skills with expert-level progressive disclosure. Use when building new skills, reviewing existing skills, debugging activation failures, encoding domain expertise, or designing skills for subagent consumption. Activate on "create…
  • 90project-memory
    Use when working in a repository that keeps durable project context, cross-session memory, incident history, decisions, or agent state across Claude, Codex, and later sessions.
  • 95skill-architect
    Design, create, audit, and improve Claude Agent Skills with expert-level progressive disclosure. Use when building new skills, reviewing existing skills, debugging activation failures, encoding domain expertise, designing skills for subagent consumption, or understanding…
  • 98context-audit
    [Adam''s Skills] Audits a repository''s Claude context-injection setup — CLAUDE.md, CONTEXT.md, docs/, .claude/agents/, and the per-project memory directory. Reports bloat, broken links, orphaned docs, security risks, missing rules in subagent prompts, settings that override…

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.