{M} Phases

Memory phases (P1–P19)

Feature phases for the memory stack (not product roadmap phases in docs/ROADMAP.md).

See also: {M}-Architecture · {M}-Deferred

---

Phase timeline

timeline
  title Aiko memory stack evolution
  P1 : Core SQLite + vector + FTS store
  P2-P4 : Spacing, entity importance, valence and salience
  P5-P7 : Emotion decay, dynamic novelty, journal promote
  P8-P10 : Tiered recall, spreading activation, Graph Studio
  P11-P13 : Provenance, scenes, cross-store context
  P14-P16 : Experience polish, LLM valence, human-feel recall
  P17-P19 : Session anchor, experience ops, arousal + lineage

Loading

---

Table

Phase	Theme	Implementation details	Primary user-visible effect
P1	Core store	sqlite-vec + FTS5; Harrier embed; LLM fact extract; user-scoped DB; chat inject; replace mem0/Qdrant	Aiko can remember explicit facts locally
P2	Spacing	access_day_count — distinct local days recalled (spaced-repetition signal for monthly gate)	Recalled-on-many-days facts become stickier
P3	Entity importance	Rank term from entity importance map (MEMORY_RANK_ENTITY_IMPORTANCE_WEIGHT)	Important people/topics rank higher
P4	Valence / salience	valence_tag, valence_score, salience_hit; monthly W_VALENCE blended with Paper I weights	Emotional and salient memories get better treatment
P5	Emotion imprint	FORGET_EMOTION_GAMMA; intensity by tag; dream prefers stored tags	Emotional memories decay more carefully
P6	Dynamic novelty anchors	Static archive anchors + dynamic mean of recent active vectors for monthly novelty	Novel but relevant memories survive better
P7	Journal promote	Optional promote of journal fragments before retention gate; safer day-pin delete coverage	Daily notes can become durable memories
P8	Rank / tiered recall	RRF KNN+FTS+graph; quick→wide; recency-among-relevant; score thresholds	Recall gets faster, broader, and less noisy
P9	Spreading activation	Walk entity_relations; depth/decay/min strength; extra neighbor memories + score weight	Related memories can surface through connected entities
P10	Memory Graph Studio	Galaxy UI; nodes/edges export; filters (status, valence, retain, entity)	Memory becomes inspectable and debuggable
P11	Hard provenance	Monthly delete requires LLM source_ids coverage of kept day-pins	Safer consolidation before deleting day-pins
P12	Scenes + 5-point valence	L2 scene_id / kind=scene; expand on recall; valence_score −2…+2	Episodes recall with richer context
P13	Cross-store	Attach related knowledge + experience after personal recall; Studio node types	Personal memory gains factual and experiential context
P14	Experience store polish	Experience DB RAG for agent runs; auto-relate threshold; entity boost	Agent traces become reusable context
P15	LLM valence on extract	MEMORY_VALENCE_FROM_LLM; extract can supply valence_score; Studio contrast	Better affect labels when enabled
P16	Human-feel recall	state_json; soft MEMORY_NEG_RECALL_AVOID; supersession narrative; Studio dims superseded	Less jarring recall and clearer updated facts
P17	Session dynamic anchor	Rolling mean of last-K query embeddings; rank boost toward current chat topic	Current conversation theme influences recall
P18	Experience / spreading ops	Config + wiring for experience supersede / spreading knobs	More controllable experience graph behavior
P19	Arousal + hard filter + lineage	arousal_score write + rank bonus; sticky-neg hard filter; lineage.py + Studio route	Safer affect-aware recall and auditable belief chains

---

Capability matrix

Capability area	Early phases	Middle phases	Late phases
Storage	P1 local DB and vectors	P12 scenes	P13+ cross-store graph
Ranking	P2 spacing, P3 entities	P8 RRF, P9 spreading	P17 session anchor, P19 arousal
Affect	P4 valence/salience	P5 emotion decay, P12 5-point valence	P15 LLM valence, P19 arousal/filter
Safety / audit	P1 explicit rows	P11 provenance	P16 supersession narrative, P19 lineage
UX / tooling	Chat injection	P10 Studio	P19 lineage API for future timeline

---

Dependency graph

flowchart TD
  P1[P1 Core store] --> P2[P2 Spacing]
  P1 --> P3[P3 Entity importance]
  P1 --> P4[P4 Valence / salience]
  P4 --> P5[P5 Emotion imprint]
  P2 --> P6[P6 Dynamic novelty]
  P4 --> P6
  P6 --> P7[P7 Journal promote]
  P1 --> P8[P8 Tiered recall]
  P3 --> P9[P9 Spreading activation]
  P8 --> P9
  P9 --> P10[P10 Graph Studio]
  P7 --> P11[P11 Hard provenance]
  P4 --> P12[P12 Scenes + 5-point valence]
  P10 --> P13[P13 Cross-store]
  P13 --> P14[P14 Experience polish]
  P12 --> P15[P15 LLM valence]
  P15 --> P16[P16 Human-feel recall]
  P8 --> P17[P17 Session anchor]
  P14 --> P18[P18 Experience ops]
  P16 --> P19[P19 Arousal + lineage]
  P17 --> P19

Loading

---

Review checklist by phase family

flowchart LR
  A[Schema changes] --> B[Config flags]
  B --> C[Ranking behavior]
  C --> D[Studio/debug visibility]
  D --> E[Retention and delete safety]
  E --> F[Paper/wiki update]

Loading

Before merging a new memory phase	Ask
Schema	Is the migration additive and safe for existing DBs?
Config	Can flags restore prior behavior?
Recall	Does ranking change only intended entry points?
Safety	Are negative/sensitive memories filtered at recall, not destroyed unexpectedly?
Studio	Can a developer inspect the new state?
Docs	Does {M}-Papers need an update?

---

Back: Home · {M}-Architecture