pulse-memory

name

pulse-memory

description

Resolve conflicts and contradictions in notes, memory, and harness state using admissions rules (archive vs memory). Always stamp context-hint tags (source, time, status, kind). Pulse sparse snippets instead of dumping transcripts. Use when claims disagree, a story was completed under pressure, the user says pulse / admissions / context tags / archive-not-memory, or runs /pulse-memory. Essays: captain.kingsparrow.space/focus/memory-issue and /archive-not-memory.

pulse-memory

Load rule: This file is the filter plane. Expand references/usecases.md for modes / thinking_path / harness_path / as_of. Expand references/admissions.md only if a tag or conflict rule is still ambiguous. Store: references/store.md. Do not paste the essays.

Archive  : residue (logs, tickets, embeddings, full transcripts)
Memory   : a past trace kept only because it is expected to change the next action
Pulse    : sparse retrieval → surgical snippet (not a dump)
Hint     : source · time · status · kind  // required on every claim you write or inject
Status   ∈ { user-stated, tool-verified, inferred, unknown }
Kind     ∈ { data, context, fact, memory }

Axioms

1. The archive may be large. What enters the live path stays sparse, dated, sourced, incomplete-ok.
2. Admissions test: would this trace change the next action enough to justify bringing it back? If no → archive, do not inject.
3. Never average conflicting traces into one fluent story. Tag both; prefer later + higher status; else not in the evidence.
4. Every sentence you promote to memory must carry hint tags. Untagged write = hypothesis, not memory.
5. Abstention is legal. If you cannot tag it, do not save it.

Sources (do not copy bodies): Pulse instead of dump · Archive is not memory

---

When to use

• Two notes, memories, or tool results contradict
• User asks to tag, pulse, admit, or clean a dump
• Writing session/career/memory files that will be retrieved later
• /pulse-memory (turn) · /workflow pulse-memory (mode: both | thinking | harness)

Skip: one-file typo with no claims about the past. Former name: archive-not-memory (alias skill redirects here).

Workflow args (file modes): as_of always (ISO-8601). thinking_path if mode is both/thinking. harness_path if both/harness. next_action recommended. Examples: references/usecases.md.

---

Always stamp (hint tags)

hint:
  source: <path | url | user | tool:name | imagined | told-later>
  time:   <ISO-8601 or "unknown">
  status: user-stated | tool-verified | inferred | unknown
  kind:   data | context | fact | memory
  lock:   open | locked

Status rank: tool-verified > user-stated > inferred > unknown.
Freshness over similarity. A later user-stated revocation beats an older similar “fact.”

---

Resolve conflicts / contradictions

1. Extract each claim. Untagged → tag now (status: unknown if unknown).
2. Same subject, incompatible predicates → conflict set, do not merge.
3. Compare time then status. Winner must still pass admissions for this next action.
4. If tied or evidence missing: not in the evidence; keep both in archive.
5. Never promote an inferred fill to locked.

Pulse shape (body ≤ ~120 tokens):

pulse:
  key:     <retrieval key>
  verify:  <hash or "none">
  prefetch:<next likely key or "none">
  snippet: <surgical text>
  hint:    { source, time, status, kind, lock }

---

Done when

• Every injected or written claim has hint tags
• Conflicts listed, not averaged
• Snippets are pulses; full logs stay archive
• “Not in the evidence” used when the filter fails

Do not: dump transcripts; invent completeness; save fluent lies as locked truth.