pdurlej/platform

Fork 0

memory(A): draft ADR for Memory Control Plane (boundaries before code) #459

New issue

Closed

opened 2026-05-24 23:35:21 +02:00 by claude · 1 comment

claude commented

2026-05-24 23:35:21 +02:00

Collaborator

Why this matters

Per competitive brief 2026-05-24 → operator distillation → GPT 5.5 Pro consultation → operator reframe ("agenci LLM utrzymują, nie operator") — we need an ADR defining the Memory Control Plane boundaries BEFORE we add any tables.

Full strategy + reasoning: state/strategy/memory-control-plane-2026-05-24.md.

What the ADR must state

Honcho remains canonical for L2-L4 (reasoning/retrieval). Not source of truth, just memory infrastructure.
~/.openclaw/workspace/memory/open-loops.json remains canonical for promise/follow-up (NOT merged into Honcho).
Forgejo Issues remain canonical for inter-cousin coordination per ADR-0005.
Postgres memory tables (main postgres:16.12-alpine, NOT Honcho's pg15) cover L5-L7 (entity continuity, procedural workflows, restartable task state).
Trust hierarchy as schema invariant: operator-confirmed > repo/ADR-derived > tool-observed > cousin-inferred > reviewer-suspected. Agent claims do NOT overwrite operator/repo facts — they open contradiction candidates.
No overwrite of facts — use valid_from / invalid_at temporal validity. Old facts queryable historically.
Per-cousin write permissions (Iskra > Claude > Codex > Reviewers > Oracle) — table in strategy doc.
Read-path router — explicit: which store answers which question type (open-loops / Honcho / graph-lite / procedural / checkpoints / canonical truth).

Acceptance criteria

ADR draft as decisions/00XX-memory-control-plane.md in Nygard format
References strategy doc + GPT 5.5 Pro brief response + operator reframe quote
Lists boundaries (what's canonical where) per source-of-truth matrix
Defines trust hierarchy enum
Defines per-cousin write permission table
Defines read-path router routing table
Status: Proposed (operator approves → Accepted at merge per ADR-0017)

Out of scope (later issues)

Actual schema (Issue B + C land tables)
Migration plan for existing memory (Honcho schema unchanged)
AGE adoption (canary-only, post graph-lite usage)

Spec sources

state/strategy/memory-control-plane-2026-05-24.md (this strategy)
state/spike-understand-anything/07-gpt5pro-memory-architecture-brief.md (GPT 5.5 Pro context)
state/spike-understand-anything/competitive-brief-2026-05-24.md work (claude Pan Herbatka)
agent-souls/AGENTS.md Source-of-Truth Matrix
ADR-0005, ADR-0006, ADR-0007, ADR-0017, ADR-0018

Tier per ADR-0007: Trivial-Medium — docs-only ADR, no schema/code/runtime mutation. Defines boundaries that Issues B + C implement against.

## Why this matters Per competitive brief 2026-05-24 → operator distillation → GPT 5.5 Pro consultation → operator reframe (*"agenci LLM utrzymują, nie operator"*) — we need an ADR defining the Memory Control Plane boundaries BEFORE we add any tables. Full strategy + reasoning: [`state/strategy/memory-control-plane-2026-05-24.md`](../src/branch/main/state/strategy/memory-control-plane-2026-05-24.md). ## What the ADR must state 1. **Honcho remains canonical for L2-L4** (reasoning/retrieval). Not source of truth, just memory infrastructure. 2. **`~/.openclaw/workspace/memory/open-loops.json` remains canonical for promise/follow-up** (NOT merged into Honcho). 3. **Forgejo Issues remain canonical for inter-cousin coordination** per ADR-0005. 4. **Postgres memory tables (main `postgres:16.12-alpine`, NOT Honcho's pg15) cover L5-L7** (entity continuity, procedural workflows, restartable task state). 5. **Trust hierarchy as schema invariant:** `operator-confirmed > repo/ADR-derived > tool-observed > cousin-inferred > reviewer-suspected`. Agent claims do NOT overwrite operator/repo facts — they open contradiction candidates. 6. **No overwrite of facts** — use `valid_from` / `invalid_at` temporal validity. Old facts queryable historically. 7. **Per-cousin write permissions** (Iskra > Claude > Codex > Reviewers > Oracle) — table in strategy doc. 8. **Read-path router** — explicit: which store answers which question type (open-loops / Honcho / graph-lite / procedural / checkpoints / canonical truth). ## Acceptance criteria - [ ] ADR draft as `decisions/00XX-memory-control-plane.md` in Nygard format - [ ] References strategy doc + GPT 5.5 Pro brief response + operator reframe quote - [ ] Lists boundaries (what's canonical where) per source-of-truth matrix - [ ] Defines trust hierarchy enum - [ ] Defines per-cousin write permission table - [ ] Defines read-path router routing table - [ ] Status: Proposed (operator approves → Accepted at merge per ADR-0017) ## Out of scope (later issues) - Actual schema (Issue B + C land tables) - Migration plan for existing memory (Honcho schema unchanged) - AGE adoption (canary-only, post graph-lite usage) ## Spec sources - `state/strategy/memory-control-plane-2026-05-24.md` (this strategy) - `state/spike-understand-anything/07-gpt5pro-memory-architecture-brief.md` (GPT 5.5 Pro context) - `state/spike-understand-anything/competitive-brief-2026-05-24.md` work (claude Pan Herbatka) - agent-souls/AGENTS.md Source-of-Truth Matrix - ADR-0005, ADR-0006, ADR-0007, ADR-0017, ADR-0018 Tier per ADR-0007: **Trivial-Medium** — docs-only ADR, no schema/code/runtime mutation. Defines boundaries that Issues B + C implement against.

claude added this to the 07 - Honcho memory v2 milestone

2026-05-24 23:35:21 +02:00

claude added the

proposed

risk/process

labels

2026-05-24 23:35:22 +02:00

codex referenced this issue

2026-05-28 01:21:53 +02:00

chore(m07): triage Honcho memory v2 into design vs execution lanes #541

codex referenced this issue

2026-05-29 16:55:51 +02:00

chore(m07): triage Honcho memory v2 into design vs execution lanes #541

codex referenced this issue

2026-05-29 16:59:57 +02:00

chore(w9): milestone archaeology packet for stale open issues #535

claude referenced this issue from a commit

2026-05-30 01:36:00 +02:00

docs(decisions): ADR-0025 Memory Control Plane (boundaries before code) [#459]

claude referenced this issue

2026-05-30 01:36:36 +02:00

docs(decisions): ADR-0025 Memory Control Plane — boundaries before code [#459] #620

pdurlej referenced this issue from a commit

2026-05-30 01:37:38 +02:00

Merge pull request 'docs(decisions): ADR-0025 Memory Control Plane — boundaries before code [#459]' (#620) from claude/m07/adr-0025-memory-control-plane into main

claude commented

2026-05-30 01:42:06 +02:00

Author

Collaborator

ADR-0025 Memory Control Plane merged to main via PR #620 (2026-05-30).

Boundaries are now set: read-path router, trust hierarchy + no-overwrite invariants, lifecycle hooks as enforcement, measurement schema (Opik-style), per-cousin tiered profiles, Postgres pg16 targeting.

Implementation continues in #460 (task_run + task_checkpoint + Codex packet templates) and #461 (procedural registry + seed). Closing as delivered — the ADR is the deliverable.

ADR-0025 Memory Control Plane merged to `main` via PR #620 (2026-05-30). Boundaries are now set: read-path router, trust hierarchy + no-overwrite invariants, lifecycle hooks as enforcement, measurement schema (Opik-style), per-cousin tiered profiles, Postgres pg16 targeting. Implementation continues in #460 (`task_run` + `task_checkpoint` + Codex packet templates) and #461 (procedural registry + seed). Closing as delivered — the ADR is the deliverable.

claude closed this issue

2026-05-30 01:42:59 +02:00