Reactions as training signals

Most agent products treat the chat surface as input and the model as black-box. To “improve” the agent, you go somewhere else (a labelling tool, a feedback form, a model retraining pipeline) and feed signals in.

Thoth’s bet: the chat surface IS the labelling surface.

Five emoji reactions on the bot’s replies feed five distinct signals into Thoth’s memory + skill system. No separate tool. The training loop happens in the place you’re already paying attention.

The five reactions

Emoji	Signal	What it does
✅	“this was right”	Mark turn verified=success. Boosts skill-compile signal in next reflection.
❌	“this was wrong”	Mark turn verified=failure. Next reflection writes a stronger what_didnt note.
🧠	“remember this”	Save the turn verbatim into MEMORY.md (with secret redaction).
🗑️	”outdated”	Mark this episode outdated. Excluded from future recall.
👤	“feedback about me”	Record as a user-feedback observation in Honcho.

Each one writes to a different layer of Thoth’s 5-layer memory architecture. They are NOT redundant.

✅ — Verified success

When you react with ✅:

1. The episode in bridge.db gets verified_status = 'success', verified_by = <your-user-id>, verified_at = <now>.
2. At session-end reflection, Thoth’s claude -p reflection subprocess sees this episode tagged success.
3. If the same pattern appears across multiple ✅-tagged episodes, reflection proposes a skill (should_skill: true in the structured JSON).
4. The skill draft writer creates .claude/skills/<slug>/SKILL.md and DMs you an approval card.

You’re effectively saying: “This is the kind of thing I want crystallized into a reusable skill.”

❌ — Verified failure

When you react with ❌:

1. The episode gets verified_status = 'failure'.
2. At reflection, the what_didnt field gets a stronger weighting.
3. The next time a similar query appears, episodic recall surfaces this failed example to the agent’s context, with the failure tag visible.
4. The agent learns not to do this again by example, not by instruction.

This is the negative training signal. Use it when the bot’s reply was actively wrong — not just suboptimal.

Don't ❌ casual mistakes

For minor bugs (“typo in summary”, “missed one Sentry issue”), correct the bot inline rather than ❌. Save ❌ for genuinely wrong replies — confusing logic, hallucinated facts, contraindicated advice. ❌ is durable; corrections are ephemeral.

🧠 — Save to memory verbatim

When you react with 🧠:

1. The bot’s reply text is appended to your MEMORY.md file verbatim.
2. Before writing, the redaction pass scans for secrets (Slack tokens, GitHub PATs, API keys, env-style *_SECRET patterns). Matches are replaced with [REDACTED:<kind>].
3. The next time MEMORY.md loads (next session boot, or after persona-drift detection), this content is in context.

Use 🧠 when the bot has surfaced a durable insight — a fact, a process, a decision — that you want it to remember next time.

Examples of good 🧠 candidates:

• “The production orchestrator is service-orchestrator-v2, not v1”
• “Cherry-picks from staging to main; never git merge”
• “EU AI Act classifies our use case as high-risk Annex III Cat 3a”

Example of a bad 🧠 candidate:

• “The Sentry dashboard is at sentry.io/orgs/{your-org}” — that’s in TOOLS.md already

🗑️ — Outdated

When you react with 🗑️:

1. The episode gets outdated = 1 in the SQLite store.
2. Episodic recall queries filter on outdated = 0. The episode is now invisible to future recall.
3. The episode still exists in storage (you can /recall it explicitly via slug if you remember it). But it doesn’t surface in routine recalls.

Use 🗑️ when an episode contains information that became wrong over time — not because the bot was wrong then, but because reality moved.

Examples:

• A reply about an API endpoint that’s since been deprecated
• A reply about a process that’s been replaced with a new one
• A reply about a person who’s no longer in the role

👤 — User feedback observation

When you react with 👤:

1. The reaction context (the bot’s reply + your reaction) is forwarded to Honcho’s identity layer as an Thoth-authored observation about you (the reactor).
2. Honcho integrates this into your peer model.
3. Future Dialectic queries about you may surface this observation when relevant.

Use 👤 when the bot’s reply revealed something about you as a user that should inform future interactions.

Examples:

• A reply asks “do you want me to summarize?” and you 👤 to signal “yes, this is the kind of question worth checking in on”
• A reply uses German in a context where you’d prefer English; you 👤 to encode the preference

Why reactions, not a labelling UI?

Reactions are:

• Always visible — every Slack thread is a candidate
• Zero-friction — one tap on mobile, one click on desktop
• Already familiar — Slack users react constantly anyway
• Multi-user — different people in a thread can react with different signals
• Idempotent — adding then removing then re-adding produces the same signal

A labelling tool would be more “powerful” in features — but the features go unused. The friction kills the feedback loop.

How reactions interact with reflection

At session-end (after /done or 30-min idle), the reflection subprocess sees the full transcript including all reactions.

The reflection prompt includes:

Of the [N] turns in this session:
- [k1] were marked ✅ verified=success
- [k2] were marked ❌ verified=failure
- [k3] were marked 🧠 saved to memory
- [k4] were marked 🗑️ outdated

Pay particular attention to ✅ patterns — these are signals to
crystallize into skills.

The reflection’s structured JSON output then gets fan-out to four writers (memory, skill drafts, persona observations, Honcho updates) per Memory L5.

Reactions are the input that drives this growth loop.

Multi-user reactions

If multiple users are in a Slack thread, each reaction is attributed to the reactor:

• ✅ from User A and ❌ from User B both register; the reflection sees them as conflicting signals and weights accordingly
• 🧠 from any user appends the reply to the shared MEMORY.md
• 👤 reactions are scoped to each individual reactor’s Honcho peer model

What reactions DON’T do

• They don’t fine-tune the model. (Thoth doesn’t fine-tune anything.)
• They don’t immediately change the bot’s behavior. (Effect is visible at next session start, after reflection has run.)
• They don’t propagate retroactively. (A 🧠 on a reply from yesterday saves that reply to MEMORY.md, but past responses to similar queries aren’t rewritten.)

What’s next

• Memory architecture — how reactions feed each layer
• Skills — how ✅ patterns become skills
• Persona stack — what MEMORY.md looks like