Brain runtime integration

Anda Bot 0.13 embeds Brain 0.12. The ordinary Formation/Recall/Maintenance loop works without a runtime configuration. Durable inbox actions, independent observations, semantic attention, utility, trust and learning are separately configured capabilities. Compilation, installed bindings, enabled automation, mechanism tests and measured business improvement are different states.

Start with ordinary memory

Run anda memory guide for an offline guide, or anda memory to check a running daemon. These commands do not call a model; the guide does not create a home or identity. The browser dashboard now opens Memory with an example, connection status, records and source quotes. Explore relationships preserves the graph view. TUI /memory and bare /brain show the same overview; explicit /brain status keeps its technical meaning.

Tell Anda a useful preference, wait for background processing, then use /new and ask about it. A reply saying “remembered” is not a storage receipt. Message badges and /memory activity report the specific conversation's processing; a completed Formation need not contain the fact you expected. Unknown acceptance is reconciled without resubmission. Source quotes link to the original message only when the durable mapping and content digests agree.

The owner can search, correct, stop using or delete supported records. Search is an explicit model operation with a complete Recall packet, warnings and default limits of 4,096 packet tokens and 32,768 planner input tokens; it has no automatic retry or total monetary cap. Corrections and removal show a ten-minute, revision-bound preview before confirmation. A correction retains the old claim and creates a new claim; undo previews another conditional correction. An interrupted change is checked by its operation ID, not submitted as a new operation.

Stopping use archives the previewed claims, cited inputs and recorded dependents. Deletion purges that bounded native set. Both persist source exclusions, fence in-flight processing and reset native/Bot Notes and processing history. Related source conversations stop contributing automatically, including their continuation chain. Other independent records, original Bot chats/files/logs/backups, already delivered model context and provider copies remain. Minimal source identifiers/digests are retained to prevent replay. Source gaps, protected dependents, retention holds or an overlarge closure refuse the operation. Restoring an old database backup without current exclusion metadata is outside this guarantee.

For a fresh conversation, use the browser composer's Memory mode control, or:

anda agent run --memory-mode no-store --prompt 'Help with this task without saving long-term memory'
anda agent run --memory-mode off --prompt 'Help using only this new conversation'

These modes constrain host-managed Brain/Notes, including implicit profile initialization and automatic Formation. Restricted modes reject unsupported nested agents, memory actions, bookmarks and cron changes. They do not make chat history, files, logs or provider processing incognito. Changing mode creates a new conversation without copying old context; existing conversations keep their persisted policy.

Product API and development dependency

/daemon/memory/v1 is an owner control plane using Bot ToolResponse, separate from native KIP and model tools. HTTP and WebSocket requests revalidate the original credential. Overview, records, search, change prepare/commit/status/discard, inbox setup and record subscriptions share the same service. Conversation activity additionally supports an authenticated caller's own specified conversation. Other records and counts are not exposed to that caller. Identifiers in product responses are strings.

This checkout temporarily patches anda_brain to sibling ../anda-brain/anda_brain for the new native product contracts, as explicitly authorized for development. The sibling package currently declares version 0.12.1; registry availability and contract parity have not been verified for this change. After the native release, update the registry constraint/lockfile, remove only the Brain patch, and re-run the contract matrix in implementation status. DB/KIP/Core/Engine remain one registry type identity. Do not restore a pre-change database as a software rollback without preserving current source exclusions.

Set up questions only when needed

Ordinary memory works without an inbox. Select Set up a memory inbox in Memory, or run anda memory inbox setup to preview the owner-only configuration. Confirm the shown digest with --apply; the wizard makes private backups, checks the original configuration digest and writes a minimal managed runtime file. Environment overrides and custom runtime files are preserved and require an explicit manual merge. Configuration has no effect until you run anda restart; verify access afterward. A config rollback does not revoke native grants.

After setup, Ask me when this memory changes creates an actual recipient-owned Watch. The controller receives an archive grant scoped only to that new Watch so cancellation needs no broad management rights. Cancellation archives the native Watch; retry/restart does not re-arm it or restore a revoked cancellation grant. Existing delivered questions are not retracted by cancelling future observation. Questions and delivery receipts are different native inbox items.

Use /memory inbox, /memory next, /memory answer <number> <text> and /memory retry <number> in the TUI. The terminal and browser save the original response/event key before sending. A lost response can be retried with the same data; changed pending text is rejected. An acknowledgement means the answer was accepted, not that a business task succeeded.

Startup

In ~/.anda/config.yaml (or your selected home):

brain:
  runtime_config: brain-runtime.yaml

Paths are relative to the configuration directory. BRAIN_RUNTIME_CONFIG overrides this field; an absolute path is accepted. Omit the setting for the existing memory behavior. Restart after changes. Runtime configuration is trusted deployment data, never a model argument, HTTP body or recalled memory.

Copy the inbox example, replace its owner CWT subject with the actual identity, and keep the configuration outside the repository. Brain validates the complete native configuration before loading any Space. Invalid adapters, unavailable environment secrets and conflicting principals fail startup. Bootstrap does not restore revoked grants on restart, re-arm Watches, grant manage_trust, or authorize Skill execution. A controller, recipient, observer and governor have distinct roles. A wildcard Bot CWT is not automatically any of these native principals.

The existing Brain background lifecycle discovers persistent work, runs one owner per shard and drains native writes at shutdown. Learning HTTP services must already be listening when configured: startup probes executor capabilities and observer identity before the Bot gateway starts. Do not configure a startup callback to the not-yet-listening gateway itself.

Inbox and identities

Existing browser tokens exported before Brain audience support must be re-exported with anda browser token and replaced in the extension settings. New tokens target anda_bot; their subject, scope and bounded expiry are retained.

Open Attention inbox in the browser Brain view. Use /brain inbox, /brain status or /brain next <cursor> in the TUI. /brain answer <id> <event_key> <answer> sends a clarification; /brain statement <id> <event_key> <statement> records an attributed statement. Use /brain help for the command forms.

Model tools brain_attention, brain_respond, and brain_runtime_status are available through tools_select. The response tool's strict schema requires answer and statement, with the unused field set to null; it sends Brain's native tagged response after validation. External IM users remain excluded from these tools. HTTP and WS use the caller's own bearer. Engine tools map the authenticated caller to the explicitly configured native principal. Anonymous/local/public access never bypasses runtime authentication. Recipient queues and audit inventories are not shared across users.

Reading a page does not claim work. complete ends a snapshot page walk, not a task. On an expired or invalid cursor, refresh from the first page. Response URLs use the returned hex id, not wake/v1/.... The browser saves each pending response before sending and retries the same event/text after a lost acknowledgement or reload. The product TUI outbox preserves that event key and text; the advanced /brain answer command still requires the caller to supply them. Same-key changed text is a conflict. A clarification answer is data, not authority; a statement is not an independent Outcome.

attention_inbox_v1 uses Brain's native persistent delivery and Attempt identity. Its native records remain the work truth after restart. Bot stores only off-graph conversation/channel/recipient/route associations; (reply_target, thread) is retained. No existing IM adapter is enabled as an automatic Brain action: send -> Result<()> cannot prove delivery or authoritative NotStarted. Unsupported adapter names fail startup; ordinary IM chat continues to use its existing permission and routing rules. No Watch is converted to unattended cron or a full-access goal.

Recall and Formation receipts

Brain HTTP KIP bodies use the application contract (command, or operations with execution, plus parameters, read, and top-level dry_run). Responses remain KIP 2.0 envelopes. Both response status levels are checked. The native Rust envelope is converted at one boundary; unsupported native request metadata, preconditions, deadlines and extensions are rejected rather than dropped.

recall_memory accepts optional budget with the pinned Brain tokenizer, max_tokens and context_tokens. Null keeps legacy behavior. Complete packets are preserved, including insufficient coverage; no additional memory is injected through artifacts or receipts. Failed recalls are marked as errors and their measured direct/nested usage is retained without summing nested totals twice. Missing billing information remains incomplete.

A host journal under bot-brain/v1/ in the existing DB object store retains Recall delivery identity, Brain conversation/receipt, Bot conversation/turn and an unambiguous model tool-call ID. Parallel identical calls remain unassigned rather than guessed. Receipts prove delivery only. Ordinary retrieval does not write utility, trust or execution authority; legacy trace receipts do not imply contribution.

Formation submission windows are recorded before sending. Accepted Brain conversation IDs are retained, and /brain formation <id> reads that specific conversation's submitted/working/completed/failed state. The global formation high-water mark is not a completion receipt. A rejected request can retry with backoff; unknown acceptance after transport loss or process interruption blocks that window for reconciliation instead of blindly resending. There is no server-side submission idempotency key, so this integration does not claim exactly-once Formation.

Independent observations and optional calibration

The native /outcomes route and typed Rust client accept independently authenticated instruments. Ordinary model tools and completion hooks cannot submit Outcomes. Native Action context/Decision records carry actual used_refs, applied_revisions and Recall receipts; the observer's OutcomeInput.utility can carry the native single-contribution witness. Delivery, use and contribution remain distinct. Unattributable ordinary tasks stay audit-only, without fabricated trials or score updates.

Startup semantic, utility and trust fields go directly through Brain's validators and native runtimes. Their model/endpoint/tokenizer/identity/calibration contracts are pinned by deployment configuration. Utility remains Concept-scoped; trust requires binary factual evidence and exact actor/predicate/context. Proposals and governance remain native Rust capabilities, with no model trust setter and no bootstrap manage_trust. Automatic application needs explicit calibrated configuration. Unknown measurements are not zero, and mechanism fixtures do not establish useful learning.

Learning and MIB

learning = ["anda_brain/learning"] is independent of mib. Build with --features learning to install workflow_http_v1 through the same runtime configuration. The learning example is a template with all automatic switches and calibration approval off. Replace identities, endpoints, pins and calibration material before deployment; its thresholds are illustrative, not calibrated defaults.

The compiled native adapter supports only tool_workflow.precondition.v1: real inspect/prepare/commit, isolated reset, stable requests, current fences/deadlines, status lookup, cancellation, complete logs and measured budgets. It probes separate executor and observer services and obtains frozen cohorts from the registered source. It uses Brain's existing Decision→Attempt→independent Outcome→Trial/Evaluation controller, archival, review and safety recovery. It does not turn arbitrary Bot shell or IM tools into a learning executor.

The MIB host still advertises persistent/no_memory, learning=false, independent_observer=false, and incomplete accounting. Enabling the Cargo feature does not silently change those experimental conditions. Normal/ungated provider trials require their real business services, independent instrumentation, full accounting, approved frozen plans and an explicit evaluation budget. No production credentials, notifications, paid model trials or cross-repository MIB changes are supplied by this integration. See MIB host and Brain's business contract.