02.9: Conditioning

Reward/punish reinforcement memory — what the bot has been thanked for and what it has been corrected on.

File: src/utils/text/context/templates.ts:280-352

Mission

When reward and/or punish conditioning is enabled for the persona, load recent grouped conditioning entries from conditioningMemoryRepository and emit them as a system-role context item describing past behaviors — “do this again” vs. “avoid this.” Each line names the action, the users who triggered the conditioning, the reason text, and a count suffix when repeated.

Input

• client, guildId
• serverId: number — the internal server primary key (not the Discord ID)
• personaLineageId: number — must be ≥ 0
• botName
• personalMemoriesEnabled (passed to convertMentions)
• rewardEnabled: boolean, punishEnabled: boolean
• convertMentions

Output

Promise<StructuredContextItem | null> — null if neither reward nor punish has any visible groups. Otherwise one system-role item tagged KNOWLEDGE_SERVER_CONDITIONING.

Content shape:

## Rewarded Behaviors
Here are past things {botName} did that got rewarded for. Strive to do them again:
- [{botName} was thanked by <@111>, <@222>. Reason: `helpful with debugging`] (3 times)
- [{botName} was praised by <@333>. Reason: `kept the joke going` with `said something punny`]

## Punished Behaviors
Here are past things {botName} did that got punished for. Avoid doing them again:
- [{botName} was corrected by <@444>. Reason: `gave wrong info about pricing`]

Side effects

• DB read (per type) — conditioningMemoryRepository.loadGroupsForPersona(serverId, personaLineageId, type) for "reward" and/or "punish".
• Past-participle resolution — getConditioningContextPastParticiple(type, actionKey) resolves the verb form (“thanked,” “corrected,” etc.) for each action kind.
• Mention conversion — assembled text passes through convertMentions; user mentions are rendered with display names.

Invariants

After this stage runs:

• Returns null if both reward and punish are disabled, or if both produce zero visible groups.
• Each enabled type contributes at most CONDITIONING_CONTEXT_MAX_GROUPS_PER_TYPE groups (see src/utils/conditioning/conditioning.ts).
• Groups with empty reasonText are filtered out (no information value).
• Skipped upstream in nativeBuilder.ts when: isUserImpersonation is true, tomoriState.persona_lineage_id < 0, or neither reward_conditioning_enabled nor punish_conditioning_enabled is set.

Configuration

Source	Field	Effect
tomoriState	reward_conditioning_enabled	Show “Rewarded Behaviors” section
tomoriState	punish_conditioning_enabled	Show “Punished Behaviors” section
tomoriState	persona_lineage_id	Scopes groups to this lineage
Constant	CONDITIONING_CONTEXT_MAX_GROUPS_PER_TYPE	Per-type cap

Extension points

Surface	Plugin-relevance
Action verb table (getConditioningContextPastParticiple)	Localized via the conditioning helper; a plugin adding a new action kind extends the verb map there.
conditioningMemoryRepository	The repository is the seam — a plugin adding a new reinforcement-type (e.g. neutral observations) would extend it.
Grouping algorithm	Currently groups by actionKey + userDiscIds + optional actionText; the grouping logic lives in the repository, not this contributor.
Section heading text	Hardcoded English here; localization would extend. → plugin plan candidate if locale-aware conditioning is needed.

This contributor is a conditioning-specific instance of the “memory type” plugin category (alongside stages 03 and 07). A plugin replacing or extending it would do so via the memory-type registration mechanism (when that exists).

Related docs

• Conditioning system: → no dedicated doc; src/utils/conditioning/conditioning.ts helper only
• Conditioning tools (reward/punish actions): tool registry (→ tool-loop pipeline)
• Server memories (parallel): 03-server-memories.md