Memory and skills
Persistent memory, AGENTS.md files, and reusable skills for TalkOps, including creation, discovery, and invocation
There are two primary ways to customize how TalkOps (talkops) works in your environment:
- Memory:
AGENTS.mdfiles and persistent memory entries that carry across sessions. Use memory for coding style, infrastructure naming conventions, and project architecture. - Skills: Reusable, on-demand
SKILL.mddirectories that the agent discovers at startup and reads only when relevant. Use skills for task-specific workflows, cloud provider best practices, and reference checklists.
Use /remember inside any session to prompt the agent to extract conventions from the current conversation and save them to memory or skills.
Want to define parameterized /command workflows with inline shell execution? See Custom commands.
Memory
Automatic convention learning
As you collaborate with TalkOps, you can teach it infrastructure and coding conventions that persist across future sessions:
> We use AWS KMS customer-managed keys for all S3 buckets and enforce bucket-owner-enforced object ownership
Or explicitly save a convention with /remember:
/remember Our S3 buckets always require KMS CMK encryption and bucket-owner-enforced ownership
TalkOps stores the convention and applies it automatically when you ask it to create or review resources in later sessions.
Where memories are stored
| Scope | Path | When it loads |
|---|---|---|
| Project (higher priority) | .talkops/memory/ | Whenever running inside that project |
| User (fallback) | ~/.talkops/memory/ | Every session across all directories |
When both scopes contain a memory with the same key, the project-level entry takes precedence. To keep memory from consuming excessive context-window capacity, TalkOps caps injected memory at 200 lines (or 25 KB) per session.
Manage memory (/memory)
Use the /memory command inside an interactive session to inspect or edit stored entries:
/memory # List all active memories
/memory save <key> # Save a new memory entry
/memory delete <key> # Remove a stored memory entry
AGENTS.md files
AGENTS.md files provide persistent instructions that are always loaded into the system prompt at session start:
- Project:
.talkops/AGENTS.mdorAGENTS.mdin the repository root—loaded whenever TalkOps runs inside that project (commit this file to Git to share conventions with your team). - Global:
~/.talkops/{agent}/AGENTS.md—loaded on every session for that agent across all repositories.
When to use global vs. project AGENTS.md
Use a global AGENTS.md (~/.talkops/{agent}/AGENTS.md) for:
- Personal communication style and formatting preferences
- Universal DevOps habits (for example, preferring
tofuoverterraformor always running dry-run diffs first) - Tool usage patterns that apply across every repository
Use a project AGENTS.md (.talkops/AGENTS.md or AGENTS.md) for:
- Repository architecture, directory layout, and environment naming (
dev,staging,prod) - Terraform state backend conventions, required resource tags, and module structure
- Kubernetes cluster context rules, Helm chart conventions, and CI/CD test commands
Skills
Skills package domain expertise—such as workflows, best practices, helper scripts, and reference docs—into reusable directories that follow the Agent Skills specification.
To preserve context-window capacity, TalkOps uses progressive disclosure: at startup, it reads only the name and description from each SKILL.md frontmatter. When a task matches a skill's description, the agent loads the full SKILL.md instructions on demand. Discovery runs again whenever you execute /reload.
Add a skill
1. Create a skill directory
Scaffold a new skill from the shell or create a directory containing SKILL.md:
talkops skills create my-eks-runbook
Each skill directory follows this structure:
my-eks-runbook/
├── SKILL.md # Required — instructions with YAML frontmatter
├── scripts/ # Optional — helper scripts executed by the agent
├── references/ # Optional — domain documentation and cheat sheets
└── assets/ # Optional — templates, examples, and schemas
2. Edit SKILL.md
Add YAML frontmatter (name and description) followed by Markdown instructions:
---
name: kubernetes
description: "Author, audit, and troubleshoot Kubernetes manifests following Pod Security Standards"
---
# Kubernetes Engineering Skill
When authoring or debugging manifests:
1. Always define explicit CPU and memory requests and limits.
2. Configure readinessProbe and livenessProbe on all long-running workloads.
3. Enforce non-root securityContext (runAsNonRoot: true).
4. Include standard labels: app.kubernetes.io/name, instance, version.
Skill discovery and precedence
TalkOps discovers skills across seven locations. When two skills share the same name, the higher-priority source overrides the lower one:
| Priority | Source | Location |
|---|---|---|
| 1 (highest) | Project Agent Skills | .agents/skills/ |
| 2 | Project TalkOps Skills | .talkops/skills/ |
| 3 | User Agent Skills | ~/.agents/skills/ |
| 4 | User TalkOps Skills | ~/.talkops/skills/ |
| 5 | Marketplace Plugin Skills | Installed plugins (~/.talkops/plugins/) |
| 6 | Subagent Skills | Bundled subagent packages |
| 7 (lowest) | Global Built-in Skills | Bundled with TalkOps (cloud-core, docker, kubernetes, remember) |
Use .agents/skills/ for portable skills shared across multiple AI coding assistants, and .talkops/skills/ for skills tailored to TalkOps workflows.
To protect your machine from untrusted files in cloned repositories, TalkOps tracks project skill trust in ~/.talkops/.state/skill_trust.json and prompts for confirmation the first time a project skill is discovered.
Invoke or pre-load a skill
Skills activate automatically when a prompt matches their description, or you can invoke them explicitly:
/skills # Browse all loaded skills in the TUI
/skill kubernetes # Explicitly invoke a skill mid-session
To pre-load a skill at launch from the shell:
talkops -s kubernetes
talkops -s kubernetes -p "Audit all deployments in k8s/prod/" -q
List and search skills from the shell
# List all discovered skills
talkops skills list
# Inspect a specific skill
talkops skills info kubernetes
# Search skills by keyword
talkops skills find terraform