Agent Knowledge And Persistence

Voxos agents are durable. Their identity, memories, skills, task plans, sessions, runs, schedules, transcripts, usage, tool events, grants, action-level tool route preferences, audits, reflections, and proposals are persisted in Postgres.

Tool Route Preferences

Each agent can keep an automatic or pinned provider preference for a semantic action such as web.search. Preferences do not grant tools. Resolution starts from the agent’s allowed tools, attaches trusted route metadata from the MCP preset catalog, and selects at most one implementation for each action. The result is copied into the effective policy for a text run, scheduled run, subagent, or voice session.

Automatic preferences use catalog priority and can move to the next available authorized implementation. Pinned preferences identify one exact tool source and server and fail closed if it is unavailable. Subagents resolve again after the parent and target policy intersection, so routing cannot restore a tool removed by inheritance rules.

Agent Identity

Each user has a default agent. The Agent page can update the display name and description. The backend stores identity in user_agents and scopes ownership by the durable local Voxos user id.

Memories

Memories are retained facts, preferences, instructions, relationships, project context, or task context. A memory has:

• Kind.
• Status such as active, proposed, archived, superseded, rejected, or hidden.
• Sensitivity.
• Source.
• Text.
• Normalized key.
• Confidence.
• Optional evidence session, transcript, or reflection ids.

Users can create, edit, and confirm memories in Knowledge. Granted agent tools can inspect or archive memories and submit suggested knowledge through the account improvement policy. Model-supplied confirmation and source claims do not establish owner approval.

Skills

Skills are reusable instruction documents. Developer skills can ship with the app, and user skills are user-owned records. Each skill has a name, slug, category, description, Markdown body, status, scope, and timestamps.

Agents can list relevant active skills and read a specific skill. Granted mutating tools can suggest new skills, stage revisions, or archive user skills. Revisions leave the active document intact until owner approval. Drafts are labelled as awaiting review; active skills remain subordinate to developer instructions and tool grants.

Curated MCP example skills

Voxos has three distinct skill sources:

• Developer skills ship with the application and are globally available.
• User skills are editable, tenant-owned database records.
• Curated MCP examples live under examples/skills/ in the repository and become ordinary user skills only when an operator selectively installs them.

List the available example groups without connecting to the database:

make example-skills-list

Preflight one group for an explicit local Voxos user before writing anything:

make example-skills-load USER_ID=<local-voxos-user-id> MCP=openbnb DRY_RUN=1

Remove DRY_RUN=1 to install the ready examples. MCP accepts a canonical group such as github, openbnb, google-workspace:gmail, or a comma-separated selection. Bare google-workspace expands to every curated Workspace product group and reports that expansion. There is no implicit all-groups mode.

Installation checks the selected user’s current MCP records and reports created, existing, blocked, and reconnect-warning results. It does not connect OAuth, enable tools, or grant them to an agent. Installed copies can be edited in Knowledge. A repeated load never overwrites an existing slug in any status, including an archived or user-edited copy. To intentionally load a newer template, permanently delete the existing user skill in Knowledge and run the same selective command again.

Task Plans

Task plans persist multi-step work. A plan stores the objective, priority, status, acceptance criteria, step list, and events. Steps can be proposed, active, in progress, blocked, completed, cancelled, or archived.

The Knowledge page edits task plans directly. The Operations page shows runs and schedules that execute work over time.

Subagents

Subagents are child agent runs. They are persisted as child AgentRunRecord rows under the same local user. Subagents keep parent, root, and session lineage and receive an effective tool policy intersected from the parent run and target agent. The runtime must use SubAgentService so depth, fanout, ownership, and tool policy checks are not bypassed.

Account History And Improvements

Authenticated owner text and LiveKit voice conversations can use search_agent_history, read_agent_history, and record_agent_improvement when granted. History includes all agents in the same account. Search defaults to seven days, accepts windows up to 31 days, and filters by channel, agent, category, or tool. Pages contain at most 100 records and use a stable cursor. Citations identify retained transcripts, tools, runs, sessions, or runtime errors. Reads add bounded context. Coverage and truncation describe the returned records; event counts are not incident totals.

History returns sanitized excerpts and operational summaries. Prompt snapshots, raw arguments, credentials, and raw runtime payloads are excluded. Phone speech is available only when the persisted call explicitly enables transcripts; private speech is withheld while operational outcomes remain visible. Phone participants and subagents cannot use the account history or learning tools.

Open Knowledge > Improvements (/knowledge/improvements) to inspect proposals, policy reasons, model assessments, citations, and previous/applied content. Approve or reject a proposal, undo an applied change, or load older improvements. Review binds to the exact saved content and detects later target edits. Undo preserves later owner changes and reports a conflict when restoration would overwrite them.

The account starts in review-only mode. Enable Automatically apply well-supported, low-risk new memories and skills to save eligible additive knowledge. Both the candidate and a separate assessment must score at least 0.95. Complete supporting evidence and current active knowledge must be available, with no detected conflict. This is a policy threshold, not a measured accuracy guarantee.

Automatic memory eligibility is limited to low-risk facts, preferences, projects, and task context supported by an authenticated owner statement or independent observations. New skills need corroboration across conversations and observed successful tool outcomes. Sensitive content, revisions, uncertain claims, and incomplete context require review. Unsafe content is rejected. Diagnostic follow-ups remain review records. Repeated submission returns the same proposal and does not reactivate rejected or undone knowledge.

Deleting a source chat or redacting its run withdraws dependent proposals and scrubs their copied content. Unchanged automatic knowledge is archived and scrubbed, including previously undone or archived targets. Explicitly approved knowledge and later owner-authored edits are retained. History retrieval results are not copied into durable tool-output evidence.

See Operations Runs And Schedules for error-triggered analysis and model prerequisites.

What Is Not User Knowledge

The product documentation index is global product documentation. It is not tenant-owned user data and must not contain provider keys, chat transcripts, private memories, phone numbers, OAuth tokens, or other user secrets.