{M} Papers

Papers and design notes

How the running memory stack relates to Paper I (cited in config/memory.yaml) and what to update.

Related: {M}-Architecture · {M}-Deferred

If Paper I lives outside the public repo (private draft / OSF), treat this page as a revision checklist, not a claim of a public arXiv URL.

---

What the code cites

From config/memory.yaml:

# Paper I §5.2 weights (salience/novelty/spacing/connectivity), with W_VALENCE blended in.
MONTHLY_CONSOLIDATION_W_SALIENCE: "0.30"
MONTHLY_CONSOLIDATION_W_NOVELTY: "0.25"
MONTHLY_CONSOLIDATION_W_SPACING: "0.20"
MONTHLY_CONSOLIDATION_W_CONNECTIVITY: "0.25"
MONTHLY_CONSOLIDATION_W_VALENCE: "0.10"

Also implemented in the spirit of that note:

• Dynamic anchors (P6) — novelty vs static [YYYY-MM] archives + dynamic mean of recent active vectors.
• Emotion-modified decay (P5) — effective half-life stretched by valence intensity.
• Soft monthly threshold — retention gate (~0.44).

---

Concept → code map

flowchart LR
  Paper[Paper I §5.2] --> Salience[Salience]
  Paper --> Novelty[Novelty]
  Paper --> Spacing[Spacing]
  Paper --> Connectivity[Connectivity]
  Paper --> Valence[Valence intensity]
  Salience --> Code1[`salience_hit`]
  Novelty --> Code2[Static + dynamic anchors]
  Spacing --> Code3[`access_day_count`]
  Connectivity --> Code4[`entity_relations`]
  Valence --> Code5[`valence_score` + decay]
  Code1 --> Gate[Monthly consolidation]
  Code2 --> Gate
  Code3 --> Gate
  Code4 --> Gate
  Code5 --> Gate

Loading

Paper I concept (as used in config)	Code	Phase
Salience	salience_hit; monthly W_S	P4+
Novelty (static + dynamic)	Monthly novelty	P6
Spacing	access_day_count	P2, P6
Connectivity	Entity graph / importance	P3, P9
Valence intensity	valence_score / tag; forget + W_V	P4–P15
Consolidation / archive	Monthly facts; day-pin policy	P7, P11
Explicit external store	sqlite-vec	P1

In code but not yet in Paper I (as cited)

Concept	Phase	Paper impact
Arousal axis + rank bonus	P19	Add dual-axis affect model and clarify abs(arousal) ranking
Hard neg filter (unsolicited suppress)	P19	Describe as recall policy, not retention/deletion
Soft neg avoid	P16	Explain score lowering before hard filtering
Session query-mean anchor	P17	Separate online chat-topic anchoring from monthly novelty
Supersession lineage API	P19	Add belief-revision audit trail
Cross-store knowledge / experience	P13+	Explain multi-store context without over-claiming one unified brain
Spreading activation	P9	Connect entity graph to candidate expansion

---

Do papers need an update?

Yes, if Paper I is the system-of-record description. The live system is ahead on dual-axis affect and recall safety policy.

flowchart TD
  A[Live implementation] --> B{Paper describes behavior?}
  B -->|Yes| C[Keep citations stable]
  B -->|No| D[Add revision note]
  D --> E{Changes retention math?}
  E -->|Yes| F[Update §5.2 weights / equations]
  E -->|No| G[Add implementation subsection]
  F --> H[Update wiki + config comments]
  G --> H

Loading

Suggested Paper I edits

1. Affect — Valence vs arousal; rank uses |a|; monthly still valence-primary until W_AROUSAL is proposed.
2. §5.2 — Document five-term blend; optional future W_AROUSAL defaulting to 0; hard filter is recall policy, not a gate input.
3. Belief revision — supersedes_id chains; lineage for audit/Studio.
4. Human-feel recall — Soft avoid vs hard filter; session anchor ≠ monthly dynamic novelty anchor.
5. Edge constraints — Explicit SQLite vs parametric long-term memory research.
6. Version table — Point at Memory-Phases (P1–P19).

---

Proposed affect-model wording

quadrantChart
  title Dual-axis affect for memory
  x-axis Unpleasant valence --> Pleasant valence
  y-axis Low activation --> High activation
  quadrant-1 Joy / excitement
  quadrant-2 Fear / anger / urgency
  quadrant-3 Sadness / low mood
  quadrant-4 Calm / comfort
  "rank bonus uses |arousal|": [0.50, 0.82]
  "negative policy uses valence": [0.18, 0.60]
  "emotion decay uses intensity": [0.32, 0.35]
  "gratitude memory": [0.82, 0.30]

Loading

Suggested concise paragraph:

The memory store represents affect with two orthogonal axes. Valence estimates pleasantness or unpleasantness and participates in emotion-modified decay and negative-recall policy. Arousal estimates activation or urgency and may provide a bounded ranking bonus by magnitude. Retention remains governed by the Paper I salience/novelty/spacing/connectivity blend plus valence until an explicit arousal retention weight is introduced.

---

Optional “Paper II” sketch

mindmap
  root((Paper II: companion memory))
    Dual-axis affect
      Valence
      Arousal
      Recall safety
    Multi-store context
      Personal memory
      Knowledge
      Experience traces
    Auditability
      Supersession
      Lineage API
      Studio visualization
    Edge constraints
      SQLite
      Quantized embeddings
      No always-on vector service

Loading

• Dual-axis affect at encode and rank.
• Companion inject safety filters.
• Multi-store memory / knowledge / experience without over-claiming one brain model.
• Auditable belief revision through supersession lineage.

---

Summary

Question	Answer
Is code ahead of Paper I?	Yes
Block engineering merge on paper rewrite?	No
Minimal paper fix	Dual-axis affect + recall policy ≠ retention gate + P19 footnote
Best next doc task	Add a P19 subsection and a small version table

---

Back: Home · {M}-Architecture