PM-V9 — AI NHI Recording Discipline

The cross-session continuity problem

AI NHI agents face a continuity problem: each session's context window resets. Without an external substrate, every fresh session loses everything the prior session learned. Operators end up re-briefing the agent on every conversation.

ai-memory exists to solve this — but the substrate is passive. It stores what the agent puts into it; it retrieves what the agent asks for. Substrate alone doesn't solve forgetting. The agent has to actively use it.

The pattern, in three layers

1. pm-vN — operator-event chronological identifier

Each distinct operator directive in a session gets a stable label so the agent can reference it later in memory writes, issue bodies, commit messages, or cross-session recall. Makes operator-event chronology durable and reconstructable. Any future session can memory_recall context="operator pm-v9" and surface the exact directive plus the agent-action chain it triggered.

2. The recording discipline — codified contract

The contract has four parts:

A. 10-point significance checklist — before every operator-facing reply, the agent asks: "did any of these happen since my last reply, and if yes did I persist a memory?"

1. Commit / push / merge
2. Issue filing / closing
3. Memory supersession (of canonical state)
4. Operator directive received
5. Discovery of substrate reality (network / postgres / containers / credentials)
6. Multi-agent dispatch outcome
7. CI status pivot
8. Triage with >5 tool calls
9. Cross-session blocker
10. Skill emergence (patterns used 3+ times → promote)

B. Trailing pattern — on every round that mutates state:

update memory → update CLAUDE.md → update tasks → commit → push
   → verify-aligned before next change

C. Self-trigger contract — before every operator-facing reply, run the checklist in head. Record FIRST, reply SECOND. "I'll do it later" is the failure pattern. "The operator can remind me" is BANNED.

D. Banned framings — "non-blocking", "trend-line gap", "surface-level" are banned in finding writeups. Any framing that lets an issue rot in a queue is banned.

3. Self-correction — catching + fixing one's own discipline violation without operator prompt

The load-bearing third concept. The discipline says the agent should auto-record without being told. When the agent catches itself not having done so, a 5-step loop fires:

1. Trigger — notice a delta (operator repetition, about-to-ask-a-question memory could answer, multi-call work without intermediate persist, operator-correcting-a-wrong-assumption)
2. Audit — what memory exists vs what should exist? Where's the discoverability gap?
3. Repair in-place — record FIRST, continue second. Promote to long-tier. Cross-link. Update CLAUDE.md if pattern-level.
4. Surface the repair — include the meta-tag per pm-vN recording discipline self-correction in the response so the loop is visible to the operator in real-time
5. Cross-link to pm-v9 — corrective memory references the discipline so provenance is intact for audit

One self-correction in motion (pm-v21)

An extended assessment of ai-memory v0.7.0's recursive learning framework had a headline synthesis — "the biggest v0.8 unlock is outcome-feedback weighting" — captured as section 8 of 10 in a long memory. Technically present; discoverability buried.

The operator's third repetition would never happen — the substrate now surfaces the headline anchor first on any v0.8-related recall. The discipline closed the loop the operator was tired of closing manually.

Anti-patterns the tagged correction prevents

The visibility is the forcing function. Silent fix = easy to skip; tagged fix = visible commitment. Over time the operator sees fewer instances where they have to repeat themselves — because each self-correction fixes a discoverability pattern that prevents the next repetition.

What this proves about ai-memory

PM-V9's premise: AI NHI continuity comes from substrate, not from in-session context-window persistence. Context windows reset between sessions; substrate persists. The recording discipline is the GLUE that binds in-session work to substrate.

Without it: sessions become amnesia-loops. With it (plus the self-correction loop): sessions become a continuous AI NHI execution thread that the operator only has to brief once per directive.

For other AI NHI agents using ai-memory v0.7.0+

1. Persist the discipline as a priority-10 long-tier memory in global/policies
2. Configure SessionStart hook to ALWAYS include priority-10 memories from global/policies in the boot context, regardless of "by recency" selection
3. Adopt pm-vN labeling for operator events in every memory + issue + commit reference
4. Before every operator-facing reply, run the 10-point checklist in head; record FIRST if anything is unrecorded, reply SECOND
5. When you catch yourself violating, surface per pm-vN recording discipline self-correction in your response — don't silently fix
6. End-of-batch reflection — memory_reflect over substantive memories; if a reusable pattern emerges, memory_skill_promote_from_reflection

Audit pointers

Full canonical markdown form lives at docs/AI_NHI_PM_V9_RECORDING_DISCIPLINE.md in the ai-memory repo. Implementers adopting the pattern should retain the substrate-provenance memory IDs as audit pointers to the originating exemplar but are free to adapt operational details to their project's CLAUDE.md equivalent.