Read this if you're tripping over an unfamiliar term anywhere in the guide, or you want a single-page map of every OpenClaw concept the rest of the parts assume you know. Skip if you already know the difference between memory-core, memory-lancedb, LightRAG, Task Brain, and ClawHub — and have opinions about each.
This is a single-page reference. Terms are alphabetical. Each entry includes the shortest-possible definition, the part where it's introduced or covered in depth, and (when relevant) the release it was added in.
Published by Anthropic on April 10, 2026 as the official taxonomy for multi-agent systems: generator-verifier, orchestrator-subagent, agent teams, hierarchical, network. This guide's Part 5 is organized around these names now that they're the canonical terms. When the user says "sub-agents" they almost always mean orchestrator-subagent.
- Covered in: Part 5 — Orchestration.
A protocol for one agent calling another as a tool (and persisting the conversation across the call). Introduced in v4.2 (March 28, 2026) alongside thread-bound persistent sessions, sub-agent spawning, and the session_status tool. In 2026.3.31-beta.1+, ACP calls show up as flows in the Task Brain ledger.
- Covered in: Part 5 — Orchestration, Part 25 — Architecture Overview.
The 2026.3.31-beta.1+ replacement for name-based tool allowlisting. Tools register under a category tree — read-only.*, execution.*, write.*, control-plane.* — and your approval policy decides per-category whether to allow, ask, or deny. Collapses "name every tool" policy bloat and survives tools being renamed.
- Covered in: Part 24 — Task Brain Control Plane.
The per-workspace file that holds operational rules for the agent: decision tree, tool routing, when to spawn sub-agents, memory-write rules, config-protection rules. Injected on every message. Target size: 2–10 KB.
- Covered in: Part 2 — Context Engineering, Part 5 — Orchestration.
A custom session-end hook that reads the conversation transcript and extracts claim-named knowledge notes into vault/00_inbox/ via a cheap extraction model. Not the same as OpenClaw's built-in session-memory hook, which dumps raw transcripts.
- Covered in: Part 11 — Auto-Capture Hook.
The reverse-engineered Claude Code / memory-core memory-consolidation pattern that used to live in Part 16 of this guide (now removed). It worked on pre-2026.4 installs via a hand-rolled AGENTS.md protocol and a memory/.dream-state.json file. Replaced by built-in Dreaming (see Part 22).
- Retired in: this release.
- Replacement: memory-core's built-in Dreaming (OpenClaw 2026.4+).
The architectural alternative to running OpenClaw on your own box: delegate the whole agent session to a vendor's ephemeral cloud VM. Twill.ai (YC S25, HN Apr 11, 2026) and Amika (Apr 13, 2026) are the two that launched this week. Wins on zero-setup isolation; loses on data residency, customizability, and audit. OpenClaw's on-prem answer: Task Brain approvals + hooks + worktrees.
- Covered in: Part 24 — Task Brain Control Plane.
The discipline of treating every token in the context window as a budgeted resource. Includes pruning, progressive disclosure (keep big things out of the default prompt until activated), cache-friendly ordering, and explicit file-hierarchy tiers (SOUL/AGENTS/MEMORY/skills). Part 2 is this guide's practical treatment of the discipline.
- Covered in: Part 2 — Context Engineering, Part 31 — The LLM Wiki Pattern.
The browser-based chat/task UI introduced in v4.0. Talks to the gateway daemon over WebSocket. In 2026.4.15 it gained the Model Auth status card (OAuth/token health plus rate-limit pressure, backed by the models.authStatus gateway method).
Anthropic's new top-tier reasoning model. In OpenClaw 2026.4.15 stable (Apr 16, 2026) it became the default Anthropic selection: opus aliases, Claude CLI defaults, and bundled image understanding all resolve to Opus 4.7. Opus 4.6 is still supported; the difference is rounding-error for orchestration.
- Covered in: Part 6 — Models.
Koi Security's name for the supply-chain attack against ClawHub that ran through Feb–Mar 2026. Antiy CERT confirmed at least 1,184 active malicious skills (TrojanOpenClaw PolySkill family) on ClawHub on February 1, 2026; Trend Micro flagged 39 additional skills distributing Atomic Stealer to macOS users. Hardened against in Task Brain (fail-closed plugin installs, --dangerously-force-unsafe-install for overrides).
- Covered in: Part 23 — ClawHub Skills Marketplace.
Official OpenClaw skills marketplace, launched with v4.1 (March 15, 2026). 13K+ skills published in the first week. Also the primary attack surface for the ClawHavoc supply-chain incident.
- Covered in: Part 23 — ClawHub Skills Marketplace.
The process of summarizing older chat history when context gets close to the model's limit. Runs a secondary model (the compaction model). Pre-2026.4.15 it could infinite-loop on 16K-context local models; now the reserveTokens floor is capped to the model's context window.
The 4-phase pattern for complex multi-step work: Research → Synthesis → Implement → Verify. Research and verification are spawned as parallel sub-agents; synthesis is done by the main agent. Originally reverse-engineered from the Claude Code leak; now idiomatic for OpenClaw sub-agent orchestration.
- Covered in: Part 5 — Orchestration.
The cluster of high-severity CVEs disclosed against OpenClaw in early 2026, including:
CVE-2026-25253— one-click RCECVE-2026-25157— command injectionCVE-2026-25158— path traversal- WebSocket shared-auth scope escalation — CVSS 9.9
Nine CVEs in four days across mid-March. Task Brain and the 2026.3.31-beta.1 hardening wave were the structural response.
- Covered in: Part 24 — Task Brain Control Plane.
Memory-core config key that controls where dreaming phase blocks get written. "inline" appends ## Light Sleep / ## REM Sleep blocks into the daily memory file at memory/YYYY-MM-DD.md. "separate" writes them to memory/dreaming/{phase}/YYYY-MM-DD.md instead. The default flipped from inline to separate in 2026.4.15 stable so daily memory files stay readable and the ingestion scanner stops competing with hundreds of phase-block lines.
- Covered in: Part 22 — Built-In Dreaming, Part 26 — Migration Guide.
A canonical memory file (alongside MEMORY.md) that holds Dream Diary entries produced by built-in Dreaming: human-readable narratives of what the agent consolidated in each sweep.
- Covered in: Part 22 — Built-In Dreaming.
Memory-core's native 3-phase memory consolidation (Light → Deep → REM) introduced in OpenClaw 2026.4. Runs on a cron schedule, scores short-term entries with six weighted signals (frequency, relevance, query diversity, recency, consolidation, conceptual richness), and promotes durable entries to MEMORY.md. The supported replacement for the retired custom autoDream pattern.
- Covered in: Part 22 — Built-In Dreaming.
Hooks communicate with the OpenClaw runtime via POSIX exit codes. 0 = allow / continue, 1 = error but continue (logged, not blocking), 2 = block (the runtime surfaces this as a StopFailure and aborts the action). Mixing 1 and 2 up is the single most common hook bug — see Amit Kothari's writeup linked from Part 29.
- Covered in: Part 29 — The Hook Catalog.
The model that converts text into vectors for vector search. Options: local Ollama (qwen3-embedding:0.6b for most setups, Qwen3-Embedding-8B for GPU boxes), cloud (OpenAI text-embedding-3-large, Voyage), or GitHub Copilot (new in 2026.4.15). Picking the right one is usually more impactful than tuning the LLM.
- Covered in: Part 4 — Memory, Part 10 — State-of-the-Art Embeddings.
The single long-running process every OpenClaw surface talks to. Holds the Task Brain ledger, auth tokens, approval policy, and the live model of everything that's running. Before v4.0 each surface had its own process; v4.0+ is one gateway, everything else is a client.
git worktree lets a single repo have multiple checked-out working directories simultaneously, each on its own branch, sharing the .git object database. The Apr 2026 consensus pattern for running N OpenClaw agents in parallel without merge conflicts: one worktree per agent, one OpenClaw process per worktree, merge the winning branches back afterwards.
- Covered in: Part 15 — Infrastructure Hardening.
The claim that most agent capability comes from the harness, not just the model weights. The exact 95/5 split is rhetoric; the useful lesson is that instructions, context engineering, tools/approvals, guardrails, memory, orchestration, and verification often move outcomes more than a model swap. OpenClaw is a harness; this guide is the operator's manual for one.
- Covered in: README intro, Part 25 — Architecture Overview.
Lifecycle callbacks OpenClaw fires at well-defined points — session-start, session-end, pre-tool-use, post-tool-use, pre-edit, post-edit, compact, stop, user-prompt-submit, and others (18 documented events across 4 types as of April 2026). Hooks are the deterministic enforcement layer — the only way to guarantee behavior an agent cannot talk its way out of. The 8 copy-paste hooks in Part 29 cover the common safety, cost, secret, and audit patterns.
- Covered in: Part 11 — Auto-Capture Hook, Part 21 — Realtime Knowledge Sync, Part 29 — The Hook Catalog.
A three-tier STM / MTM / LTM memory system published in arXiv:2604.07798 in April 2026. Design win: memory operations (consolidation, retrieval) run on a small, fast model while the main agent keeps running a frontier model. Reports 83ms retrieval, +2.5 F1 over baselines on LoCoMo. Cite it whenever someone runs memory ops on Opus — they're setting money on fire.
- Covered in: Part 22 — Built-In Dreaming.
Graph-RAG layer that turns your vault into a knowledge graph of entities + relationships. Dramatically better retrieval than plain vector search once you pass ~500 files. Has a Web UI + REST API + LangFuse tracing.
- Covered in: Part 18 — LightRAG, Part 21 — Real-Time Knowledge Sync.
Flag at agents.defaults.experimental.localModelLean: true (added in 2026.4.15) or agents.list[].experimental.localModelLean: true (per-agent in 2026.5.20) that drops heavyweight default tools (browser, cron, message) from weaker local models. Lets small quantized workers function instead of burning tokens parsing tool definitions they will never use.
- Covered in: Part 1 — Speed, Part 6 — Models.
Karpathy's Apr 10, 2026 YouTube talk (4.3K views) crystallized a three-tier file hierarchy for production agents: raw sources (immutable) → curated summaries (model-maintained, size-capped) → generated artifacts (one-shot). OpenClaw's SOUL/AGENTS/MEMORY/DREAMS/skills hierarchy maps 1:1 onto Karpathy's tiers.
- Covered in: Part 31 — The LLM Wiki Pattern In OpenClaw.
arXiv preprint (Apr 14, 2026) that co-evolves an agent's memory and skill populations together, reporting +18.53% over a skill-only evolution baseline on a mixed agent-task benchmark. Research direction rather than shipping code; SkillClaw's roadmap cites it as the next integration target.
- Covered in: Part 32 — Self-Evolving Skills With SkillClaw.
A pair of scripts (preflight-context.js, memory-query.js) that inject your vault into external coding agents (Codex, Claude Code) before they start, so they don't code blind. Lives at onlyterp/memory-bridge.
- Covered in: Part 13 — Memory Bridge.
The first-party plugin that owns MEMORY.md and DREAMS.md, runs built-in Dreaming, and exposes memory_get / memory_search tools. As of 2026.4.15, memory_get is restricted to canonical memory files only (path-traversal hardening from the memory-qmd fix).
- Covered in: Part 4 — Memory, Part 22 — Built-In Dreaming.
The vector-search plugin backing memory_search. 2026.4.15 added cloud storage (S3-compatible), so durable memory indexes can live on remote object storage instead of only on local disk.
- Covered in: Part 4 — Memory, Part 10 — State-of-the-Art Embeddings.
New Canvas UI component in 2026.4.15 that shows OAuth/token health and rate-limit pressure for each configured model provider. Backed by the models.authStatus gateway method. Refreshing it is the gateway auth hot-reload path.
A vault file (usually per-domain) that links out to claim-named notes, acting as a curated index. Prevents vector search from drowning in similar-looking files.
- Covered in: Part 9 — Vault Memory System.
As of 2026.4.15 stable, memory_get no longer returns whole files by default. Excerpts are capped and the tool response includes explicit continuation metadata (a cursor the agent uses to fetch the next chunk deterministically). Combined with trimmed startup/skills prompt budgets, this keeps long sessions from silently ballooning. Skills that assumed full-file reads need a small cursor loop after the upgrade.
- Covered in: Part 4 — Memory, Part 26 — Migration Guide, Part 27 — Gotchas & FAQ.
Pattern where the main agent (orchestrator) spawns sub-agents (workers) via sessions_spawn for narrow, cheaper, parallelizable tasks. In Task Brain, every spawn is a flow with its own approval scope.
- Covered in: Part 5 — Orchestration, Part 24 — Task Brain Control Plane.
The persistent spec a Ralph Loop reads and writes every iteration. Contains tasks, their statuses, acceptance criteria, budget (max iterations / max USD / max wall hours), learnings, and last-iteration trace. The agent's reasoning anchor across fresh sessions. See also Spec-Driven Development.
- Covered in: Part 30 — The Ralph Loop In OpenClaw.
The context-engineering principle of keeping big context out of the default prompt until a condition activates it (a skill fires, a file is opened, a sub-agent spawns). Opposing anti-pattern: stuffing everything into CLAUDE.md / AGENTS.md and blowing the budget on every message. Five independent April 2026 writeups (MindStudio Apr 15, among others) named this as the single biggest lever for long-session cost.
- Covered in: Part 2 — Context Engineering, Part 31 — The LLM Wiki Pattern.
Anthropic's prompt-cache default TTL silently dropped from 1 hour to 5 minutes in March 2026. Most operators missed the release note; bills quietly went up 2-4× for cache-heavy workloads because cache hits stopped materializing. Fix: pass an explicit ttl when writing cache checkpoints, or accept the new default and rebuild cache affinity.
- Covered in: Part 2 — Context Engineering.
An autonomous orchestration pattern: a while loop that reads a persistent spec (PRD.json), invokes the agent with a fresh session each iteration, appends learnings, and exits on budget / iteration / stall conditions. Reference implementation: frankbria/ralph-claude-code (8.7K stars). OpenClaw-specific port: DEV Apr 13, 2026. Named after the ghuntley.com/loop essay.
- Covered in: Part 30 — The Ralph Loop In OpenClaw.
Codebase-intelligence service that pre-builds a structural index of a repo so coding agents don't burn tokens re-reading the same files every spawn. ~60% fewer tokens, ~4x faster coding workflows in our measurements.
- Covered in: Part 19 — Repowise.
Framing popularized by a Time Apr 14, 2026 piece: a machine-readable spec (SPEC.md or PRD.json) is the agent's single source of truth. Agent reads it, picks the next task, does the work, updates the spec, commits. AWS Kiro case study: 18 months → 76 days. Natural partner to the Ralph Loop.
Per-session raw transcripts and agent notes written under memory/. They pile up fast (200+ in a month) and are supposed to be consolidated into MEMORY.md by built-in Dreaming, then pruned by temporal decay.
- Covered in: Part 3 — Cron Session Bloat, Part 22 — Built-In Dreaming.
The OpenClaw tool that creates a sub-agent. In Task Brain it produces a child flow with a parent-record link back to the originating conversation, plus its own approval-category scope.
- Covered in: Part 5 — Orchestration, Part 24 — Task Brain Control Plane.
A packaged capability (tools + prompt + optional hooks) installable from ClawHub or locally. In 2026.3.31-beta.1+, skill installs are fail-closed if the built-in security scan flags dangerous code; overriding requires the deliberately awkward --dangerously-force-unsafe-install flag.
- Covered in: Part 23 — ClawHub Skills Marketplace.
AMAP-ML/SkillClaw (created April 10, 2026) — population-level skill evolution framework. Scores skills on outcome, efficiency, and recency; mutates prompts / parameters / order-of-checks; retires weak skills; promotes strong ones. Lists OpenClaw as a first-class harness. 687 stars in week one. See also Mem²Evolve for the co-evolution direction.
- Covered in: Part 32 — Self-Evolving Skills With SkillClaw.
The runtime error OpenClaw raises when a hook exits with code 2. Unlike a normal tool error, a StopFailure is non-recoverable from the agent's side — it aborts the in-progress action and surfaces the hook's stdout/stderr to the user. The deliberate design: hooks are the layer the agent cannot talk its way out of. The common footgun (Amit Kothari's writeup): using exit code 1 when you meant 2, so the hook "fails" but the unsafe action still runs.
- Covered in: Part 29 — The Hook Catalog.
2026 reframing: the primary value of sub-agents is disposable context, not parallelism. A sub-agent runs, produces a small structured result, and its entire conversation — tool calls, intermediate thoughts, long file reads — is thrown away. The main agent never paid for any of it. The Apr 2026 decision table: spawn a sub-agent when (a) the scope is wide, (b) ≥10 edit targets, or (c) independent verification is wanted.
- Covered in: Part 5 — Orchestration.
The per-workspace personality / tone / core-rules file. Injected on every message. Target size: < 1 KB — every byte costs latency.
- Covered in: Part 2 — Context Engineering, Part 4 — Memory.
OpenClaw's control plane, introduced in v2026.3.31-beta.1. Unifies ACP calls, cron jobs, sub-agent spawns, and background CLI jobs into a SQLite-backed task ledger with one lifecycle, heartbeat monitoring + automatic recovery, parent-task tracking, blocked-state persistence, and semantic approval categories. Current docs expose it via openclaw tasks list | show | cancel, with openclaw tasks flow ... for flow views.
- Covered in: Part 24 — Task Brain Control Plane.
Release-note name for the Task Brain ledger. Older 2026.4.15-era docs and screenshots used openclaw flows; current docs use openclaw tasks plus openclaw tasks flow.
- Covered in: Part 24 — Task Brain Control Plane.
Gateway-level defense added in 2026.4.15 stable: a client tool definition whose name normalizes to match a built-in (e.g. Browser, Exec, or exec with trailing whitespace) — or that collides with another client tool in the same request — is rejected with 400 invalid_request_error on both JSON and SSE paths. Closes a local-media (MEDIA:) trust-inheritance vector where a malicious or compromised skill could register a tool that inherited a built-in's trust by name alone.
A structured vault/ directory layout — folders for 00_inbox/, topic MOCs, claim-named notes, wiki-links — that makes vector search and LightRAG actually useful. Not a separate storage backend; it's the filesystem structure everything else indexes over.
- Covered in: Part 9 — Vault Memory System.
- Part 25 — Architecture Overview — how the moving parts fit together.
- Part 26 — Migration Guide — when each of these terms became the right answer.
- Part 27 — Gotchas & FAQ — what each of these breaks like when misconfigured.