Octobrain

Permanently delete memories. Irreversible — requires confirm=true. Use memory_id for single deletion, or query+filters for bulk removal. memory_id deletes any memory remember returned, project or global; query deletes only in the given scope (omitted: the session's project), so a project query never deletes global memories. Don't forget memories just because they're old — importance decay handles that. Only delete when information is wrong; for a replaced fact, memorize the new one with supersedes instead.

Knowledge base with six commands. The 'source' parameter (when used) ALWAYS refers to a SINGLE FILE or URL — never a directory; passing a directory path is an error. URLs are fetched anonymously — public documentation and information pages only, never API endpoints that need credentials (GitHub Actions logs, private repos, etc. return 403; use the platform's own CLI for those). 'search': semantic search across indexed content — provide source (single URL or file) to auto-index on-the-fly, omit to search all indexed sources, knowledge boxes included. 'store': save raw text under a unique key (session-scoped scratch, auto-cleaned, never durable) — error if key exists, delete first to replace. 'delete': remove stored content by key. 'read': fetch and return the FULL text content of a single URL or file — use ONLY as a last resort when search results are insufficient; prefer 'search' for targeted retrieval. 'match': search indexed content by regex pattern (like grep) — returns matching lines only; prefer 'search' for semantic queries, use 'match' for exact string/regex patterns. 'sync': re-index knowledge boxes — run it after you create, edit or delete files under '<repo>/.box/'. Supported file types: .html, .txt, .md, .pdf, .docx. Knowledge boxes are the durable, shared knowledge base: files under '<repo>/.box/' (one topic per Markdown file, grouped in folders such as rules/, guides/, decisions/; the root README.md is not indexed), plus boxes subscribed with the octobrain CLI. Box results show source 'box://<box_id>/<path>'; for this project's box the file is '<repo>/.box/<path>' — edit it with your file tools, then 'sync'. A box:// source in search or match narrows to that one file; never pass a .box/ file path as source (it indexes a duplicate). One fact or decision belongs in memorize, not a box file.

Store information, insights, or context in memory. Call remember first to avoid duplicates. Set source='user_confirmed' for user-stated facts (importance 0.8-1.0), 'agent_inferred' for AI conclusions (0.3-0.6). Skip transient state or things easily re-derived. For a longer reference document, write a knowledge box file instead (see the knowledge tool). Use related_to[] to link the new memory to existing ones in the same call. Relationship types: related_to, depends_on, supersedes, similar, conflicts, implements, extends, achieves, closes. Knowledge updates: when a new fact replaces or corrects an existing memory (a value changed, a decision was reversed), memorize the new fact with related_to=[{target_id: <old_id>, relationship_type: 'supersedes'}]. Retrieval then ranks the current fact above the outdated one, which stays queryable for history. remember first to find the old_id. Goal workflow: 1. memorize a 'goal' type memory for the task — captures intent 2. For each contributing memory: memorize with related_to=[{target_id: goal_id, relationship_type: 'achieves'}] 3. When the task closes: memorize the completion / lesson-learned note with related_to=[{target_id: goal_id, relationship_type: 'closes'}] (goal_id must be a 'goal' memory with at least one Working achieves source). This triggers automatic consolidation — your closing memo becomes the consolidated parent, all Achieves sources transition to Consolidated state with dampened importance (still queryable for audit). Importance of the closing memo is bumped to max(sources) * 1.1. No separate consolidate call needed. global=true (only available when session scope is locked): stores the memory in the shared global scope instead of the locked project scope. Use this for cross-project facts — user preferences, personal habits, universal conventions, tool choices — things that apply regardless of which project is active. Do NOT use global for project-specific knowledge.

Semantic search over stored memories. Call before memorize to avoid duplicates, and at task start to load context. Results include 1-hop graph neighbors automatically. A project scope also returns global memories. Prefer 2-5 related query terms for broader coverage. Results show [CONFIRMED]/[INFERRED] trust labels. For temporal questions ("last week", "in May"), set created_after/created_before (ISO-8601) to scope the time window — you know today's date, so compute the bounds.