Cortex Memory Multiverse — Bitemporal History + Branchable Timelines
Date: 2026-06-21 · Status: Design concept (operator-directed; "a new direction but a critical one") · Owner: cortex-memory + program-review-prime
Origin: operator on G11 — "two things: the current facts AND the complete history (where I am + everything done). Unlike human life, I want the privilege to start again from a specific point — e.g. go back to London and start from there — a replay, potentially with a different outcome. A new direction but a critical one. The principle: a multitude of possibilities. A Multiverse world."
This extends the supersession/truth-maintenance gaps (G11 / G95 / G97) from "overwrite the stale fact" into "version + branch the timeline."
1. The three properties
| # | Property | Meaning | Human analogue |
|---|---|---|---|
| P1 — Current state | "where am I now" answerable in O(1) | the live fact (e.g. lives in Lisbon) | normal recall |
| P2 — Complete history | "everything that was true / done, in order" | the full version timeline (London → … → Lisbon) | autobiography |
| P3 — Branch & replay | fork from any past state into an alternate trajectory that evolves independently, potentially to a different outcome | "go back to London and start again" | impossible for humans — the privilege cortex gets |
P1+P2 = bitemporal memory (each fact carries valid-time + transaction-time; supersession appends a new version, never destroys the old). P3 = branching (a timeline/branch axis) — coexisting alternate worlds = the multiverse.
2. Why it is critical (trajectory, not feature)
Branches are counterfactual / what-if exploration of the compounding loop. "Start from London with a different action" is exactly the §0 evolutionary ideology — try a stepping-stone in a new direction, keep the elite, never lose the learning — made first-class in memory. The multiverse is the substrate for evolutionary search over life-trajectories: many possible futures branched from shared history, compared by outcome (CCMLP slope per branch). It is the memory complement to the swarm's OBSERVE→MUTATE→SELECT loop.
3. What exists today (ground truth, current main)
- History substrate EXISTS: L4 deep archive is append-only + hash-chained — the immutable spine every branch shares up to its fork point. (the foundation for P2)
- Partial bitemporal EXISTS:
fact_keys.py:266keepsfact_key_revisions(append-every-transition). (a first P1+P2 on the fact-key path only) - Branching DOES NOT EXIST: no timeline/branch axis anywhere; semantic facts are append-only with no supersession (
consolidation.py:451-571, fact_id = content sha256, no arbitration). No fork/replay primitive. (P3 is net-new)
4. Data-model impact (must be designed WITH MT-ARR)
Branching adds a branch axis to the principal tuple: {tenant, agent, user, branch}. This is why it must be co-designed with the MT-ARR storage decision — branch_id is another scope axis exactly like tenant_id:
- Copy-on-write branches (don't duplicate every store): a branch records only its divergence from the parent after the fork point; reads fall through to the parent up to the fork (git-like).
- L4 = the shared immutable trunk; branches diverge after their fork commit.
- Retrieval becomes branch-scoped (recall within a branch's lineage = parent history up to fork + branch-local writes).
- Interacts with RTBF (G31) and the L4 hash-chain tenancy decision (MT-ARR Part 3 P-3) — branches multiply the audit/erase surface.
5. Buildable now vs. chartered direction
- NOW (G11, the first slice): extend fact-key
NEW/UPDATE/CONTRADICTIONarbitration (fact_keys.py:266) to the semantic store so supersession keeps the current state AND the full history (delivers P1 + P2 on the main timeline). High value, reuses working code, no branching yet. - CHARTERED (new — G103 Memory Multiverse): the branch/fork/replay primitive (P3) — copy-on-write
branch_id, fork-at-state, branch-scoped retrieval, branch comparison by outcome. Designed with MT-ARR; likely after the storage substrate decision (B vs C) since branches and tenants share the same scope-axis machinery. Expect experiments (this is a research direction).
6. Gaps
- G11 (this slice): bitemporal supersession on the semantic store — P1+P2. KEEP high.
- G95: the truth-maintenance umbrella G11 rolls into.
- G97: retrieval-triggered reconsolidation (recall updates a fact) — a bitemporal write.
- G103 (NEW): Memory Multiverse — branchable timelines (P3). Net-new, critical new direction; co-designed with MT-ARR; copy-on-write branch axis. Registered 2026-06-21.
7. Honest status
Concept pin — P1/P2 partial (fact-keys only), P3 unbuilt (CLAUDE.md §1). Nothing branches today. G11 delivers the history-preserving supersession now; the multiverse (G103) is a chartered new direction gated on the MT-ARR storage decision (the branch axis and the tenant axis are the same machinery).