Cortex Memory — Master Index (THE cortex entry point)

Cortex Memory — Master Index (THE cortex entry point)

Created 2026-06-21 · rewritten to the ratified direction 2026-06-26.

This is the single canonical entry for the cortex memory program. It is the cortex anchor; the
program-wide priority order lives in [[delivery/PROGRAM-ROADMAP.md]] (cortex is one program within it).
Everything cortex hangs off this file.

Direction (operator-ratified, kulbir 2026-06-26). Read this first — it supersedes the 2026-06-21

scope framing below it.

0. The ratified direction (what cortex IS)

1. Cortex is THE universal memory backend. Every agent and every application that needs to recall
like a human does — Conduir teams, Claude, Codex, Antigravity, future agents — uses cortex. Anything
with history it must recall is a cortex client. There is no second memory system.
2. ONE cortex — the twin/vendored copies are debt, not architecture. Today there are 3+ physical
memory trees (the canonical cortex_memory/, the ~20K-LOC symposium/memory/ twin, a vendored
cortex-memory/src/symposium/ tree, +a 4th evolution_memory.py) held together by a byte-identity
parity gate. That gate was always only a stopgap. Single-cortex was always the plan; the twin is
retired by CON-4147 (decomposition + full gap register:
[[delivery/CORTEX-SINGLE-CORTEX-CONSOLIDATION-GAPS-2026-06-26.md]]). Named blocker: CON-4203
(twin tenant-RLS) + the contracts/privacy decision.
3. One reach surface — MCP for all. P7 interfaces converge on an MCP-memory seam (CON-4202, under
P8/CON-4209) so every agent reaches the one cortex the same way; the witness-SSE stream (CON-4201) and
single write-path are the runtime proof the contract is consumable.
4. Durable + erasable — crypto-shredding reconciles RTBF with the immutable ledger. The P5 Ledger is
tamper-evident + hash-chained AND supports true GDPR erasure by hashing ciphertext and destroying the
per-user key (not the bytes). Design (swarm-hardened):
delivery/CORTEX-L4-CRYPTO-SHREDDING-DESIGN-2026-06-26.md (CON-3994 / successor CON-4235) —
dead citation, not on evolution/main, found 2026-07-28 by check_wikilinks.py (see §1 table row below).
5. Swappable provider + elastic platform are deferred successors, not blockers. Hexagonal
MemoryProvider port = CON-4221; the Postgres-native elastic-RLS platform = P8/CON-4209.
Today's SQLite-tier cortex is the substrate; these raise the ceiling later (chosen by trajectory, not
a today's-performance bake-off).

1. Live authoritative docs (consult these for current state — not memory)

2. Design principles (the "why" — forged in the 2026-06-21 review, still load-bearing)

3. The delivery epics (JIRA)

(Earlier P0–P3 umbrellas CON-3989/3990/3991/3992 fed the substrate/formation/retrieval/consolidate work
folded into the above.)

4. Honest status (CLAUDE.md §1 — documentation never closes a ticket)

Substantial P4–P7 implementation merged to main this cycle (tenant fail-closed CON-4187; tenant-scoped
reads CON-4210; cross-user recall CON-4169; twin re-single-source CON-4200/#1441; action-audit CON-4195;
approval-gate CON-4196; crystal quarantine CON-4198; witness-SSE CON-4201; grounded-analyst CON-4131; the
cortex-memory audit profile). None of the four epics is Done — they remain In Progress pending an
independent /epic-compliance PASS (executor does not self-certify, §4) and the operator-gated set:

• P5 RTBF crypto-erasure build greenlight + Vault CORTEX_L4_LEDGER_KEY provisioning (tamper-evidence inert until keyed — no flag-off theater).
• P4 live-loop go on real/replayed traces (CON-4197/4199 — unblocks the compounding thesis gate).
• P6/P7 single-cortex consolidation sequence (CON-4147) + recall root-cause under load (CON-4003/4001).

The path to PASS is: merge + close the agent-fixable children, land the operator decisions, then an
independent audit. Consult §1's live docs (not this summary) for the current per-ticket truth.