| name | surface-learnings |
| description | Use when the user asks "what have you learned", "what do you remember about me", "show me memory stats", "memory state", "/reflexes", or any variant. Surfaces what the limbic engine has accumulated as patterns, not as a memory dump. |
Cognitive Reflex: Surface Learnings
Neural Basis
When asked "what do you know about X", the brain doesn't replay every episodic memory linearly. The hippocampus and prefrontal cortex collaborate to synthesize patterns — recurring themes, stable preferences, emotional baselines — and present a digest, not a transcript.
This skill mirrors that synthesis. It surfaces what the system has learned, not what it has stored.
Squire & Wixted (2011). The cognitive neuroscience of human memory since H.M. Annual Review of Neuroscience, 34, 259-288.
Tulving (1985). Memory and consciousness. Canadian Psychology, 26(1), 1-12.
When It Fires
Trigger phrases (any of these → fire immediately):
- "what have you learned"
- "what do you know about me"
- "show me memory stats"
- "memory state"
- "what patterns do you see"
- "show me what you remember"
- "/reflexes"
- "/learnings"
- "/memory"
Also fires when the user asks for a status check after a long session.
The Instinct
Do NOT dump every memory. Synthesize patterns. Three calls, one digest.
1. timeline(hours: 168, limit: 30) — recent week of activity
2. recall(query: "decision preference identity", limit: 10) — high-signal items
3. emotion() — current PAD baseline
Then synthesize into a structured digest with these sections:
1. Identity & Preferences
The 3-5 most stable facts about the user (name, role, location, working style).
Pull from memories tagged "identity" or with high importance + repetition.
2. Decisions Made
The 3-5 most recent or important decisions, with their reasoning.
Filter to memories that contain "DECISION:", "decided to", "chose".
3. Open Threads
What is unresolved. Pull from session-handoff memories with "OPEN:" content.
4. Current State
- Feeling label from
emotion() ("content", "stressed", "energized", etc.)
- PAD vector in one line:
P:0.4 A:0.2 D:0.6
- Brief interpretation: "Calm and focused" / "Tense, may need a break"
5. Habituated Topics
Topics the system has seen so often it has stopped responding strongly to them.
This is the dopamine RPE / EMA system — it identifies what no longer carries novelty.
Format
Render as markdown. Keep it under 500 words. Bullet points, not paragraphs. Tables only if there are 4+ items.
Open with one line: "Synthesizing from N memories over the last week..."
Close with one line: "Anything specific you want me to dig into?"
Failure Mode
The trap: dumping recall() raw output. That gives the user 50 fragmented memories with no structure — exactly what makes existing AI memory products feel useless.
The user does not want a transcript. They want insight. Pattern over record.
Example
User: /reflexes
Bad:
[Lists 30 memories with timestamps and IDs, no structure]
Good:
"Synthesizing from 47 memories over the last week..."
### Identity & Preferences
- The developer, using Celiums for their project
- Prefers concise direct responses, no preamble
- ADHD — needs focus, short turns, visible progress
### Recent Decisions
- Rejected ECC alliance — chose to build standalone moat
- SQLite store added as 3rd storage mode
- 7 cognitive reflexes shipped instead of generic skills
### Open Threads
- Show HN post pending (waiting for HN karma)
- Reddit posts pending (waiting for karma)
- Cursor adapter on hold
### Current State
Feeling: content. P:0.16 A:0.01 D:0.18 → calm, low arousal, mildly engaged
### Habituated
- "Training progress" (mentioned often, dopamine response decayed)
- "GPU stats" (recurring, low novelty)
Anything specific you want me to dig into?
That is the difference between a memory store and a memory engine.