Helixir Memory Charter

DRAFT v0.1 — awaiting owner approval. The write path currently runs in "defer, don't destroy" mode (increment 2, #34): destructive verdicts on the conflicts below are DEFERRED — both facts are stored, the dispute lives on a charter_deferred CONTRADICTS edge, and resolve_contradiction settles it (retract executes the supersede then, with history). Non-destructive conflicts are surfaced to the agent in add_memory.needs_clarification, but every decision still executes. Blocking semantics activate only after this charter is approved.

This charter governs what Helixir may decide on its own when writing memories, what it must escalate to the agent (and through the agent — to the human), and what it must never do. Three layers, strongest first.

1. Constitution (immutable — changed only by explicit human edit)

These rules are not available to charter self-learning and override everything below.

• C1. Never auto-delete. Enforced in code: a DELETE verdict from the decision engine is executed as SUPERSEDE — the old fact stays in history with the delete-intent recorded in the supersession reason, and the conflict is escalated. Memory is an elder brain: it forgets nothing silently. (The library-level delete() remains as an explicit administrative action; it is deliberately not exposed over MCP.)
• C2. Never overwrite memories marked immutable (system seeds, approved charter rules, memories the user marked final).
• C3. Preferences, goals and opinions are never rewritten silently. Any CONTRADICT, and any UPDATE/SUPERSEDE touching these types — even at high engine confidence — is escalated. A reversed preference may be a real change of mind, a different project context, or an extraction error — only the human knows which.
• C4. raw_input memories are never modified or superseded. They are the source of truth that survives extraction mistakes.
• C5. Low-confidence destructive operations escalate. UPDATE / SUPERSEDE with decision confidence below 70 is flagged for review.

2. Learned rules (grown from precedents, each with provenance)

Rules appear here when the user explicitly answers "allow always" to a clarification, or approves an agent proposal after repeated identical answers. Every rule links (BECAUSE edges) to the episodes that created it.

(empty — no precedents yet)

3. Defaults (thresholds; tunable in config)

• Cosine ≥ 0.98 against an existing memory → exact duplicate → NOOP, silent.
• Cosine < 0.70 against everything → genuinely new → ADD, silent.
• CONTRADICT / CROSS_CONTRADICT decisions → execute (both facts are kept, linked by a CONTRADICTS edge — non-destructive) and flag in needs_clarification.