Eve library

OSHUN V1 In-Memory Runtime State Classification

Spec

Revision 1

On this page

Status: Active Owner: OSHUN Engineering Last Updated: March 23, 2026 Related: docs/releases/v1/audits/cross-cutting/in/memory-runtime-state.md, docs/releases/v1/specs/web-app-admin-v1-prd.md, docs/releases/v1/specs/studio-review-and-approval-workflows-dependency-and-data-contract-map.md, docs/releases/v1/specs/studio-background-jobs-and-progress-ux-dependency-and-data-contract-map.md

1. Purpose#

This document is the authoritative V1 classification for V1-ARCH-048: assign every audited in-memory runtime-state usage to one of three buckets:

  • acceptable cache
  • transitional prototype
  • release blocker

The source inventory is docs/releases/v1/audits/cross-cutting/in/memory-runtime-state.md.

2. Classification Rubric#

2.1 Acceptable Cache#

Use this label when all of the following are true:

  • the in-memory state is not the system of record
  • losing it on restart does not lose user data, review state, or job state
  • the value can be safely recomputed, rehydrated, or repopulated from durable systems

2.2 Transitional Prototype#

Use this label when any of the following are true:

  • the state is meaningful product state, but the implementation is clearly prototype-oriented and not yet the canonical V1 path
  • a durable implementation exists, but memory mode is still exposed as an explicit development or fallback option
  • restart loss is undesirable, but the state is not one of the PRD's named production-complete durability gates

2.3 Release Blocker#

Use this label when any of the following are true:

  • the state is explicitly covered by the PRD durability requirement in section 16.4
  • losing the state breaks review, approval, assistant continuity, research continuity, or other trust-sensitive user-facing flows
  • the in-memory store currently acts as the only practical system of record for a quality-critical runtime

3. Summary#

Classification totals from the audited inventory:

  • acceptable cache: 2 usages
  • transitional prototype: 10 usages
  • release blocker: 12 usages

The release-blocking findings cluster around four themes:

  • canonical assistant memory and session state
  • review and approval state
  • research workflow state
  • canonical substrate services still booting on memory-backed stores

4. Classification Matrix#

Usage Classification Rationale Primary Follow-On
apps/oshun/bff/src/notifications/preferences-store.ts transitional prototype User-visible preference state should persist, but it is not one of the PRD's named production-complete blocker classes. V1-ARCH-049
apps/oshun/bff/src/notifications/read-state-store.ts transitional prototype Read/unread notification state is meaningful continuity state, but not a named release gate in PRD 16.4. V1-ARCH-049
apps/oshun/bff/src/notifications/digest.ts transitional prototype Digest scheduling and last-delivery state is user-facing and restart-sensitive, but not a canonical V1 durable system-of-record boundary. V1-ARCH-049
apps/oshun/bff/src/notifications/quiet-hours.ts transitional prototype Quiet-hours settings are product state, but they do not fall into the PRD's highest-risk blocker buckets. V1-ARCH-049
apps/oshun/bff/src/routes/device-tokens.ts transitional prototype Device token registration should be durable before production push delivery, but it is a shell adapter concern rather than the core V1 blocker set. V1-ARCH-049
apps/iris/api/src/services/memory.service.ts release blocker This is canonical assistant memory state, explicitly aligned with the PRD's durable-memory requirement and the Iris ownership boundary. V1-ARCH-049, V1-ARCH-052
apps/iris/api/src/services/conversation.service.ts release blocker Conversation history and assistant continuity are trust-sensitive and restart loss would break replay and continuity expectations. V1-ARCH-049, V1-ARCH-052
apps/iris/api/src/services/agent.service.ts release blocker Agent and task state requires continuity and auditability once exposed through the canonical assistant substrate. V1-ARCH-049, V1-ARCH-052
apps/iris/api/src/services/knowledge.service.ts release blocker User knowledge-base documents and chunks are quality-critical assistant data, not safe ephemeral cache data. V1-ARCH-049, V1-ARCH-052
services/psyche/orchestrator/src/orchestrator/session/state.py default memory-backed session state and checkpoints release blocker Assistant session state requiring replay or audit is explicitly called out by PRD 16.4. A default memory-backed state manager is not production-complete. V1-ARCH-049, V1-ARCH-052
services/psyche/orchestrator/src/orchestrator/caching/semantic_cache.py memory cache and memory vector-store modes acceptable cache This is cache state, not canonical session or job truth. Loss on restart should only reduce reuse efficiency, not corrupt user or workflow records. none
apps/sophia/knowledge-graph/src/main.ts memory-backed knowledge-graph service release blocker Sophia is the canonical grounding substrate. A primary service entrypoint that loses graph state on restart is not production-complete. V1-ARCH-049, V1-ARCH-054
apps/isis/workflow-registry/src/services/workflow.service.ts explicit in-memory backend transitional prototype The registry already has a durable PostgreSQL path. The remaining issue is that in-memory mode is still available for a quality-critical control-plane service. V1-ARCH-049
apps/isis/gpu-worker/src/workers/gpu-worker.ts worker-local callbacks and loaded-model registry acceptable cache This state is local operational cache layered on top of a durable queue and durable job store. It is not the workflow source-of-truth. none
apps/lilith/svc-review/src/app.ts review queues, items, decisions, assignments, and escalations release blocker Review state is explicitly named by PRD 16.4 as durability-critical. V1-ARCH-049, V1-ARCH-050, V1-ARCH-051
apps/lilith/svc-conversation memory session store, memory fallback, and in-process conversation state release blocker Live assistant session continuity and replay-sensitive state cannot depend on process memory or silent memory fallback. V1-ARCH-049, V1-ARCH-052
apps/lilith/svc-spiritual-guidance/src/conversation-memory.ts session, journey, and memory stores release blocker This is assistant continuity and user-journey state, not a disposable cache. Restart loss breaks the intended user experience. V1-ARCH-049, V1-ARCH-052
apps/lilith/bff/src/utils/long-term-memory.ts release blocker Long-term assistant memory held in the BFF bypasses the canonical Iris durability boundary and loses state on restart. V1-ARCH-049, V1-ARCH-052
apps/lilith/bff/src/utils/user-preferences.ts transitional prototype User preferences should persist, but this module is not itself one of the PRD's named release-gate state classes. V1-ARCH-049
apps/lilith/bff/src/utils/translation-service.ts transitional prototype Translation lifecycle state is meaningful product data, but the more urgent release gate is the linked review and approval workflow state. V1-ARCH-049
apps/lilith/bff/src/utils/translation-review-workflow.ts release blocker This module directly holds review and approval workflow state in memory. V1-ARCH-049, V1-ARCH-050, V1-ARCH-051
apps/lilith/bff/src/utils/translation-open-research.ts release blocker Research submission, review, approval, and publication state is explicitly in the PRD's durability focus area. V1-ARCH-049, V1-ARCH-054
apps/lilith/bff/src/synthesis-engine.ts pendingJobs transitional prototype The module is a mock generation path and not the canonical durable Isis generation control plane, but it should not remain a production-facing job store. V1-ARCH-049, V1-ARCH-053
apps/lilith/bff/src/api/webhooks/webhook-service.ts default in-memory storage and delivery queue transitional prototype Delivery continuity matters, but the audited implementation reads like a development-grade integration surface rather than the canonical V1 trust boundary. V1-ARCH-049

5. Consequences#

The classification means the V1 sequence should proceed as follows:

  1. V1-ARCH-049 must create explicit replacement plans for every release blocker listed above.
  2. V1-ARCH-050 and V1-ARCH-051 should concentrate on the review and approval blockers in Lilith review surfaces.
  3. V1-ARCH-052 should concentrate on the Iris, Psyche, Lilith conversation, Lilith spiritual-guidance, and Lilith BFF long-term-memory blockers.
  4. V1-ARCH-053 should ensure that any remaining production-facing generation job state routes through the durable Isis path.
  5. V1-ARCH-054 should ensure that research workflow state is durably persisted in the research-bearing services.

6. Notes on Exclusions#

The following audited services did not need a runtime-state classification entry because the inspected in-memory structures were not cross-request service-state stores:

  • apps/sophia/search-api
  • apps/isis/generation-api
  • apps/isis/output-registry

Those services remain relevant to later durability work, but they did not add new classified in-memory runtime-state findings in this pass.