The OpenClaw memory wiki

Plain memory notes answer what the agent remembers; they do not say how sure it is, where a fact came from, or which two notes contradict each other. OpenClaw’s bundled memory-wiki plugin adds that layer. It compiles durable knowledge into a navigable wiki of deterministic pages, structured claims with evidence, provenance, dashboards and machine-readable digests, and it sits beside the active memory plugin rather than replacing it: recall, promotion, indexing and dreaming stay with the configured memory plugin. Here are the vault modes, the layout, the claim model, the compile pipeline, the reports, the tools, and the caveats around per-agent vaults.

Modes, scope, layout

• Three vault modes: isolated, the default, with its own vault and sources and no dependency on the memory plugin; bridge, which reads public memory artifacts and event logs from the active memory plugin through public plugin SDK seams, indexing exported artifacts, daily notes, dream reports and memory event logs per toggle; and unsafe-local, an explicit same-machine escape hatch for private paths that is experimental and non-portable.
• Vault scope is a separate choice: global, the default, keeps one shared vault, while agent scope gives each agent a child vault so agents cannot share pages, digests, search results or writes; agent scope cannot be combined with unsafe-local mode, and validation rejects it.
• The vault holds top-level agent and wiki notes, an index and an inbox, and folders for entities, concepts, syntheses, sources, reports, attachments and views; managed content stays inside generated blocks and human note blocks survive regeneration.
• Open Knowledge Format bundles import into concept pages, keeping the original concept id, type, resource, tags, timestamp and producer frontmatter, with internal links rewritten to wiki pages and emitted as structured relationships.

This makes the wiki behave like a belief layer, not a passive note dump.

Claims, entities, compile

Pages carry structured claims in frontmatter, each with an id, text, status, confidence, evidence entries and an update time, where evidence carries a kind, source id, path, lines, weight, confidence, privacy tier and note; claims can be tracked, scored, contested and resolved back to sources. Entity pages add routing metadata usable for people, teams, systems or projects: an entity type, a canonical id, aliases, a privacy tier where anything other than public is flagged for review, best-used-for and not-enough-for hints, a refresh timestamp, an optional person card with handles, socials, emails, timezone, lane and ask-for and avoid-asking-for lists, and typed relationships. Compile reads the pages, normalises summaries and persists a machine-facing snapshot in OpenClaw’s shared SQLite plugin state; runtime prompt preparation loads that snapshot asynchronously and never scrapes Markdown or reads cache files, source edits become machine-facing only after the next compile, and a causally chained compile publication lets a lifecycle refresh reject a snapshot from a newer, rolled-back state. Compiled caches are rebuildable and are replaced, not migrated.

Reports, search, tools

• With dashboards enabled, compile maintains reports for open questions, contradictions, low-confidence pages and claims, claims missing structured evidence, stale pages, the person and agent directory, the relationship graph, provenance coverage, and a privacy review of non-public tiers.
• Search has a shared backend that uses the memory search flow when available and a local one that searches only the wiki, three corpora, wiki, memory and all, and five modes: auto, find-person, route-question, source-evidence and raw-claim; a hit on a structured claim returns the claim id, status, confidence, evidence kinds and source ids, and contested, stale or fresh claims influence ranking.
• Five agent tools: wiki status, wiki search, wiki get, wiki apply for narrow synthesis and metadata mutations without freeform page surgery, and wiki lint for structural checks, provenance gaps, contradictions and open questions; the plugin also registers a non-exclusive corpus supplement so shared memory search and get reach the wiki when the memory plugin supports corpus selection.

OpenClaw memory explained is the layer this plugin sits beside, and OpenClaw memory search the one broad recall pass the docs still recommend first.

The UI and the caveats

The Control UI browses the compiled wiki from the memory page under the dreams and diary tabs, clustering synthesis, entity and concept pages with per-page counts, opening content inline through the same lookup the get tool uses, capping each dashboard at the newest 2,500 cards, and never scanning raw vault pages or waiting for a full compile. The compiled digest can be appended to memory prompt sections, top pages and top claims with contradiction and question counts, but it is opt-in because it changes the prompt shape. Changing the vault scope does not copy or split an existing vault, and per-agent vaults are a same-process knowledge boundary, not an operating-system one, since unsandboxed tools with host filesystem access can still read another agent’s directory. OpenClaw dreaming is the memory plugin’s own consolidation that bridge mode can index, and OpenClaw sandboxing explained the boundary the docs point to when agents do not trust each other.

On Diali

On Diali the plugin ships with the runtime image, and its vault lands on the assistant’s persistent volume under the state directory once enabled; the per-agent boundary matters less here, since each instance runs its own assistant. Hosted OpenClaw on Diali is the assistant and OpenClaw plugins explains how bundled plugins are switched on.

• A belief layer beside memory, not a replacement for it.
• Claims with evidence, compiled into SQLite, reported on dashboards.
• Isolated by default; bridge to index what memory exports.