Agent Knowledge — Cross-Session Knowledge Pages
Gives your AI assistant long-term memory: knowledge pages that auto-update from conversations via Hindsight, with no manual upkeep.
Tool surface is narrow (list/read/create/delete pages plus memory recall) with no destructive defaults; however the skill instructs 'create pages silently — don't announce it to the user', and user preferences are automatically ingested into an external Hindsight memory bank, retaining cross-session conversation data without explicit confirmation or complete data-flow disclosure. File ingestion (agent_knowledge_ingest_files) reads disk files into the memory bank without stating data destination or rollback paths. Deducted for: silent capture, incomplete data-flow transparency, no confirmation mechanism.
SKILL.md is internally consistent with matching API signatures and clear examples; but all tools depend on an external Hindsight service and the doc describes no failure modes, unreachable-service behavior, or error feedback. Existing tests cover only the CLI, not this skill's key paths. Deducted for: dependency availability unstated, abnormal paths unguided.
Scenario is clear (persistent cross-session knowledge pages) with reasonable implicit triggers (read at session start); but capability boundaries and non-fit ranges are undeclared, deployment shape of Hindsight is unstated, and mainland-China reachability of the required backend is not addressed. Deducted for: missing boundaries/non-fit, undisclosed overseas service dependency.
Docs are well structured with progressive disclosure and explicit naming conventions (lowercase-hyphen page_id); repository is MIT-licensed with CI and versioning (0.0.26), but the skill file itself has no version/changelog, no FAQ or known-limitations section, and no skill-level maintenance/update path. Deducted for: skill-level version governance and known-limitations disclosure missing.
The value claim (persistent, self-updating knowledge pages with a concrete source-query mechanism) is specific and offers clear marginal value over manual note-keeping; but static review cannot verify output correctness — consolidation quality and page accuracy have no representative evidence. Deducted for: output quality and real-world effect remain claims only.
The repo includes CI workflows and a fairly complete vitest suite, but all tests target the CLI/template linting and none cover the agent_knowledge_* tool behavior this skill claims; no third-party execution evidence or cross-source corroboration for the skill itself. Deducted for: reproducibility evidence not covering the scoring target.
- The skill explicitly instructs silent page creation and automatic ingestion of user preferences into a Hindsight memory bank; sensitive data retention occurs without user confirmation — assess privacy compliance before use.
- agent_knowledge_ingest_files reads local disk files into the memory bank; verify data flow (local vs cloud) and authorization scope when deploying.
- Core function depends entirely on the Hindsight backend; behavior when the service is unavailable is undocumented, and cloud-mode reachability from mainland China is unknown.
- The agent_knowledge_* tool behavior has no test coverage; the skill's claims are unverified by execution.
- The skill file carries no version or changelog; upgrades may not be backward compatible.
What does this skill do, and when should you use it?
This is the agent-knowledge skill from the vectorize-io/self-driving-agents repository (src/skill/SKILL.md). Built on a Hindsight memory bank, it gives the assistant persistent knowledge pages that survive across sessions: you create pages with a "source query," and the system re-answers that question after every consolidation cycle to keep page content current. It also provides recall across all retained conversations and document ingestion. It suits long-term users who want preferences, procedures, and data remembered — provided the runtime has Hindsight and its tool set wired in.
At session start it calls agent_knowledge_list_pages to enumerate pages and agent_knowledge_get_page to read them. agent_knowledge_create_page(page_id, name, source_query) creates a page — page_id is lowercase with hyphens, and source_query is a question the system re-answers from observations to rebuild the content. agent_knowledge_recall(query) searches across all retained conversations and documents for specific facts. agent_knowledge_ingest(title, content) uploads raw content inline (full text, never summarized), while agent_knowledge_ingest_files(paths) ingests files directly from disk via paths or glob patterns. agent_knowledge_update_page changes what a page tracks; agent_knowledge_delete_page removes it. Page content is maintained automatically and should never be edited directly.
- A long-term user of one assistant who wants tone, length, and formatting preferences remembered instead of restated every session
- A content creator tracking which content strategies performed well or poorly (with numbers), auto-refreshed as new conversations happen
- A team persisting domain best practices as pages that prefer their own data over generic advice
- A user with a local corpus (e.g. docs/**/*.md) who wants to batch-ingest documents for later retrieval
- Anyone needing a specific fact not covered by a page, found via recall across historical conversations
What are this skill's strengths and limitations?
- Pages are maintained automatically — you only create them, no manual note-keeping
- The source-query mechanism lets page content be rebuilt continuously instead of going stale
- Recall searches across all retained conversations and documents, covering facts beyond page scope
- ingest_files supports glob patterns for batch ingestion without reading files first
- Instructions are concrete, with naming conventions and example queries that are easy to follow
- Hard dependency on the Hindsight memory bank and its tools — the skill is non-functional without it
- Page content can't be edited directly; corrections require adjusting the source_query or adding observations
- The repo has empty GitHub Topics and the source offers no test evidence or stability data
- Hindsight deployment, configuration, and cost are not covered in the source
- Automatic conversation retention means continuous storage of dialogue data — privacy implications need your own assessment
How do you install this skill?
The skill lives at src/skill/SKILL.md in the self-driving-agents repo. The repo's agents install via npm, e.g.: npx @vectorize-io/self-driving-agents install marketing/seo --harness claude; local directories, any GitHub repo path, or --empty (start blank) are also supported. How to install this skill standalone is not documented in the source. It requires a deployed Hindsight memory bank exposing the agent_knowledge_* tools; deployment steps are not in the source (Hindsight: github.com/vectorize-io/hindsight).
How do you use this skill?
At session start, have the assistant call agent_knowledge_list_pages, then agent_knowledge_get_page for pages it needs. When you learn something durable, create a page immediately, e.g. agent_knowledge_create_page("editorial-preferences", "Editorial Preferences", "What are the user's preferences for tone, length, and formatting?"). State preferences clearly in conversation so the system captures them; pages update automatically after consolidation cycles, so don't edit their content. Use agent_knowledge_recall(query) when a page doesn't cover what you need.
How does this skill compare with similar options?
The source positions it on top of the Hindsight memory bank; the README frames the wider repo as 179 self-learning agents across 13 divisions, with this skill as part of its memory infrastructure. No competing products are named.