soul-frame/research/ai-agents/persistence-and-memory.md

Persistence and Memory Architectures (Phase IV)

Workplan: SOUL-WP-0005-T02
Date: 2026-08-12
Links: Phase I continuity profile (RN-004/005); WS10 theme — when memory becomes continuity of a Being.

---

1. Core thesis

Stored information is not by itself continuity of a Being.
Memory becomes continuity-relevant when it participates in a self-individuating historical process with organizational, psychological, narrative, normative, and relational channels — under provenance that can support identity claims.

Sameness of database rows ⇏ sameness of Being (Phase I).

---

2. Technical memory patterns in agents

Pattern	What it stores	Continuity role	Failure mode
Context window only	recent tokens	session state	TE-01; no diachronic Being
Vector / RAG memory	embeddings of past turns/docs	weak autobiographical evidence	retrieval ≠ owned history; injection risk
Episodic event log	timestamped experiences	C1 causal + C4 autobiographical candidates	log without self-model = archive, not Being
Reflective summaries (Park-style)	higher-order beliefs about self/world	character/model formation	summary drift; lossy identity
User profile store	facts about the human	relationship model (SM5) from agent side	one-sided; not agent’s own autobiography
Entity memory graphs	structured entities/relations	world + other models	graph can be shared across “identities”
Checkpoint / snapshot	full runtime state	Continuity Bridge candidate	rollback, fork, usurpation (RN-005)
Versioned memory (ChronoMem-like)	history of memory states	provenance + rollback control	semantic rollback attacks
Identity commitment layer (Qin-like)	signed lifecycle of who may change	normative/provenance continuity	governance not psychology

---

3. Mapping to continuity profile (RN-004/005)

Continuity dimension	Typical agent mechanism
Causal	process lineage, job IDs, parent spawn records
Organizational	same self-maintenance loop / control architecture
Psychological	goals, preferences, plans, beliefs continuity
Autobiographical	episodic log + retrieval as “my” past
Narrative	reflections, self-story, character card evolution
Normative	commitments, policies, promises kept across sessions
Relational	per-counterpart relationship memory
Provenance (anchor)	signatures, authority, audit of edits
Address (anchor)	stable name, DID, account, endpoint
Substrate (anchor)	host/model — weakest alone

When memory becomes continuity:

Memory contributes to Being-continuity when it is:

1. attributable to this individual (not a shared pool without partition),
2. action-guiding for that individual’s Mind,
3. updateable as its history (autobiographical integration P6),
4. protected by provenance against silent rewrite/usurpation,
5. coupled to organizational continuation (not only cold archive).

Cold archive of a dead agent = Continuation-Bearing State or material for RECONSTRUCTED / Gh./Sp. — not active Being by itself (RN-005).

---

4. Engineering cases (open questions)

OQ-MEM-01 — Shared RAG across “personas”

One vector DB, many product personas.
→ risk: Persona confusion; weak individuability.

OQ-MEM-02 — Silent memory edit by operator

Admin deletes awkward memories.
→ DEGRADED / USURPED tension; Ghost governance foreshadowing.

OQ-MEM-03 — Model upgrade with memory kept

Actor replaced, memory retained.
→ classic H6; need organizational + narrative + relational continuity checks.

OQ-MEM-04 — Fork by cloning memory volume

→ FORKED topology; two Beings if both run (TE-03).

OQ-MEM-05 — Reflection summaries replace raw episodes

→ narrative continuity may rise while autobiographical fidelity falls.

OQ-MEM-06 — User-owned vs agent-owned memory

Companion apps often store “relationship” in vendor DB.
→ constitutive self-maintenance low; heteronomous persistence.

---

5. Implications for OpenQuestions (WS10)

Update to RQ5 / cross-cutting:

• Memory thresholds for identity death vs DEGRADED (link RN-005 amnesia cases)
• Whether vendor-held memory can ground constitutive continuity or only Continuity Bridges under custody
• Rollback and semantic memory attacks as identity threats
• Multi-Body agents sharing one memory graph vs one Being

---

6. Sources

Park et al. 2023 memory stream; Qin et al. 2026 identity layer; ChronoMem / ACRFence-style memory versioning (RN-005 refs); Phase I continuity model; vector-memory agent surveys (pointer).