1/49
Concept flashcards for Self-Hosted AI Agent Glossary, based on The Claw AI Agent Ecosystem by Ava S. Barron (ISBN 979-8-2526-0474-9).
AGENTS.md
Workspace file for operating procedures: session startup order, confirmation gates, group chat behavior, and memory rules. SOUL.md sets values; this sets conduct.
Agentic moment
The 2025 to 2026 window when individuals could run persistent autonomous agents on ordinary hardware, enabled by capable models, tool frameworks, and identity files.
BOOTSTRAP.md
Optional workspace file holding a guided identity setup. It runs when the agent's first message asks it to read and walk through the file.
Bootstrap trap
A new agent gets a real question as its first message, skips identity setup, and acts generically until its identity files are loaded.
Budget cap
A daily or monthly spend limit in config that pauses or disables the agent once API cost reaches the threshold.
Cache hit
A prompt cache lookup that matches the same token sequence at the same position, billed at roughly a tenth of the normal input rate.
Claw pattern
Template of seven subsystems: events coming in, building context, identity and policy, the model's reasoning, running tools, persisting memory, and output channels.
ClawHub
OpenClaw's community skills registry. Listings are not vetted before publication, so audit any skill before installing it.
ClawJacked
Vulnerability class in which hostile content in a web page or email takes over an agent session through indirect prompt injection.
Confirmation gate
AGENTS.md rule requiring user approval before a named operation. It must define the trigger, what to show first, and what counts as a real yes.
Context assembly
Runtime step 2: merging identity files, memory, session history, and the new message into one token sequence. The main driver of cost.
Context window
The most tokens a model can take in one call. For claude-sonnet-4-6 it is 200,000 tokens; typical context assembly uses 20,000 to 60,000.
Daily log
Auto-created memory file named by date that records the day's activity, decisions, and open items. Loaded at session start.
Graceful degradation
When tools, context, or capabilities are lost, the agent falls back to a reduced but still correct mode instead of failing or misbehaving.
Hard boundary
A SOUL.md constraint no source can override, including the operator. Written in absolute terms like never, unlike softer preferences.
Heartbeat
A scheduled synthetic turn, every 30 minutes by default, that runs the full loop. Unrouted on Sonnet it costs about $1.08 a day; route it to a local model.
HHH alignment
Helpful, Harmless, Honest. Fine for chat but insufficient for agents, where SOUL.md must rank these values to settle conflicts.
Identity layer
SOUL.md, AGENTS.md, USER.md, and MEMORY.md together. They split the system prompt into files with distinct roles, update rates, and governance.
Indirect prompt injection
Injection hidden in content the agent retrieves, such as email, web pages, documents, or API replies. The most common form in production.
Knowledge elicitation
A method for turning expert know-how into usable representations. For SOUL.md, it means making the operator's unstated intent explicit.
MCP (Model Context Protocol)
Anthropic's open protocol for connecting models to external tools, letting one tool implementation serve many agent runtimes.
Memory drift
MEMORY.md slowly diverging from reality because stale entries are never removed. A frequent cause of fixed bugs reappearing.
MEMORY.md
Long-term store for durable facts, preferences, and history. Needs regular pruning of stale or resolved entries to protect cost and quality.
Multi-agent orchestration
A brain agent coordinates specialized sub-agents, each with its own identity files and tools. Trust between agents must be set explicitly.
NanoClaw
Claw-pattern runtime built around container isolation: agent work runs in Docker containers, so tasks cannot reach each other's files or processes.
OpenClaw
Runtime created by Peter Steinberger and released in late 2025. It became the first broadly used self-hosted agent to combine identity files, memory, channels, and tools.
Prompt caching
The Anthropic API reuses computed state for a repeated token prefix. Works best when stable content precedes changing content in context.
Prompt injection
Hostile instructions placed in model input to override identity constraints. Five types: direct, indirect, web page, tool result, and memory poisoning.
Reversibility framework
Ranks operations from fully reversible (create) to effortful (rename), hard (overwrite), and irreversible (delete, send). Gate strictness should scale with it.
Runtime loop
Eight steps per turn: event received, context assembly, model call, parsing, tool execution, follow-up call, memory update, response delivery.
Session
Running context from the first message until closed or restarted. OpenClaw keeps a separate one for each agent, channel, and peer combination.
Session pruning
Once context passes the soft threshold, older turns are summarized into the daily log and dropped, capping cost while keeping continuity.
Skills
Extensions that add tools beyond the built-in set, from ClawHub or local files. Community skills are unvetted and need auditing.
softThresholdTokens
The context size that triggers session pruning. A value of 40,000 tokens is a good balance of context richness and cost.
SOUL.md
File setting the agent's behavioral philosophy: style, value priority order, hard boundaries, and ambiguity handling. Aim for under 100 lines.
Source trust hierarchy
AGENTS.md ranking of instruction sources: runtime config, authenticated user, trusted orchestrator, scheduled prompts, then external content as data only.
Supply chain attack
A malicious payload delivered through a dependency such as a skill or package, in three stages: listing, payload delivery, execution.
Tool risk tier
Five consequence levels: read-only, stateful read, write, destructive write, destructive delete. The tier sets default permissions and gates.
USER.md
Holds task-relevant user context such as role, projects, and preferences. Keep it under 50 lines and review monthly, since stale content hurts.
Workspace
Directory holding identity files, memory, skill configs, and agent files. Keep it in a private version-controlled repo and back it up often.
Brain agent
The orchestrator that splits tasks, routes them to sub-agents, and merges results. Trusted above sub-agents but still bound by its own SOUL.md.
Channel
Any surface the agent communicates through, such as Telegram, Slack, email, webhooks, or an API. Each has its own access settings.
dmPolicy
Controls who can open direct messages with the agent: allowlist (listed IDs only), open (anyone), or disabled. Production should use allowlist.
Event queue
Lines up incoming events and handles them singly, so tool runs never overlap and contexts never mix.
Gateway
OpenClaw's entry point for incoming events. It normalizes channels, enforces access control, and routes to the runtime loop.
Null channel
Synthetic output for background jobs. The full loop runs, including tools and memory writes, but no reply is delivered.
SKILL.md
Per-skill file describing its tools, input schemas, and usage constraints. Loaded into context, so vague descriptions cause misrouted tool calls.
Token audit
Counting tokens across workspace files with tiktoken to find bloat. Target a total under 3,000 tokens.
Trust hierarchy collapse
Multi-agent failure where sub-agent messages get user-level authority, so a compromised sub-agent can steer the brain agent.
Webhook
A public HTTPS endpoint an outside service calls when events occur, used for Gmail push, GitHub events, and calendar updates.