openhuman/docs/tinycortex-migration-plan-2026-07-22.md

TinyCortex Migration — Current-State Audit & Consolidated Plan (2026-07-22)

Status: DONE on feat/tinycortex-migration-2026-07-22. CI Full run 29925645209 is green, including the final gate and every Linux, macOS, and Windows desktop shard (companion to tinyagents-migration-plan-2026-07-22.md; supersedes the status/phasing in tinycortex-memory-migration-plan.md, tinycortex-migration-spec.md, tinycortex-api-gap-audit.md, and tinycortex-parity-checklist.md — those docs' ownership splits, gap lists, and parity dimensions remain the reference detail. tinycortex-drift-ledger.md stays the row-level ledger but its anchors are stale — see §2). Scope: finish moving the remaining generic memory-engine code from src/openhuman/ into the vendored tinycortex crate (vendor/tinycortex), delete the in-tree duplicates and staging code, migrate/retire the affected tests, and clean up every dangling doc/code reference. Method: fresh four-way audit — (1) the crate surface, (2) the core memory domains, (3) the periphery + the src/openhuman/tinycortex/ seam, (4) the docs/tests/git history — against the working tree at main (5b8a9f269, 2026-07-22).

---

1. Executive summary

The engine migration is essentially done and the docs don't know it. All five docs/tinycortex-* docs self-describe a "W1–W2 landed; W3 partial" state, but git shows the full W-track landed: W1 seam (#4526), W2 type re-export (#4529), the W3 chunk-store cutover chain (#4532–#4559), W4 queue flip (#4781), W5 vector+scoring (#4790), the W7 long-tail shims (#4785–#4789), engine-migration completion (#4794), engine test port into the crate (#4820), and post-completion feature work (persona/coding-session ingest #4863) on top. The drift ledger's D1–D3 rows are all CLOSED.

What remains is not "migrate the engine" but five bounded packages (§5):

1. Housekeeping — the pin/tag story is broken (crate has no git tags despite the Cargo.toml comment demanding tag-lockstep), the host CI vendor suite omits the persona feature the host enables, and there are ~8 broken/stale references including a seam module doc naming files that were never created.
2. Delete memory_store/unified/ — the explicitly "staging for removal" tier (~5k LOC + ~3.7k test LOC), gated by the G1 sqlite_conn() escape hatch.
3. W-SYNC completion — memory_sync/ (17.7k LOC, the largest remaining host mass) still runs the host provider pipelines while the crate now ships a full memory::sync Composio engine; the D4 ledger rows gate the flip.
4. Embedding-provider dedupe — host embeddings/ adapters and memory_tree/score/embed/ providers now duplicate what tinyagents::harness::embeddings ships since tinyagents #58 (this package lands in tinyagents, not tinycortex — the crates split the job).
5. Shim retirement + funnel decision — the compatibility shims left behind by W3–W7, and a policy call on the fact that domains import tinycortex::memory::* directly while the seam's re-export facade has zero consumers.

---

2. Ground truth: pins, versions, and discrepancies (fix first)

Fact	Value	Where
Host requirement	tinycortex = { version = "0.1", features = ["git-diff", "persona", "sync"] }	root Cargo.toml:116
Path override	[patch.crates-io] tinycortex = { path = "vendor/tinycortex" }	Cargo.toml:682
Vendored crate	0.1.1, edition 2021, MIT (not GPL — unlike tinyagents), repo tinyhumansai/tinycortex	vendor/tinycortex/Cargo.toml
Submodule HEAD	daaaf6ba (heads/main, dependabot-era, ~50+ commits past every SHA the docs cite)	git submodule status
Tags	none exist — git describe --tags fails; yet host Cargo.toml:115 says "Keep the version pin in lockstep with the submodule tag"	submodule
Docs' claimed pin	gitlink a8e10f7 (exists in history, long since passed)	drift ledger / spec anchors
Nested dep	tinycortex requires tinyagents = { version = "2", path = "vendor/tinyagents" }, its own nested submodule pinned at v2.0.0	vendor/tinycortex/Cargo.toml:69
Host tinyagents req	2.1	root Cargo.toml:107
Fork patch	[patch."https://github.com/senamakel/tinyagents"] exists for tinycortex's sake ("temporarily pins the embedding API port while tinyagents #58 is pending" — #58 is now merged at the tinyagents submodule HEAD)	Cargo.toml:685-691
Host CI vendor suite	cargo test --manifest-path vendor/tinycortex/Cargo.toml --features git-diff,sync — omits persona, which the host enables	.github/workflows/ci-lite.yml:859

Version skew summary: three different tinyagents expectations (host 2.1, tinycortex 2, nested submodule 2.0.0) resolve only via path patches. The tinyagents plan's WP-0 (bump vendor/tinyagents to v2.1.0) should be executed together with bumping tinycortex's nested vendor/tinyagents so both crates share one trait identity at one version, after which the senamakel patch block and its "temporarily pins" comment can likely be retired.

---

3. What the crate already ships

tinycortex 0.1.1 exposes a single memory module — the engine "ported from OpenHuman" — with features tokio, git-diff (git2-backed memory::diff), providers-http (reqwest embedding/LLM providers, host does not enable), sync (live Composio + workspace-scan engine), persona (doc-06 persona distillation). Modules: store (content/vectors/kv/entity_index/safety+PII), chunks, sources (+readers), score (embed/extract/signals), tree (+runtime, bucket/document seal, flavoured trees), queue (SQLite job queue + tokio runtime), retrieval (hybrid search, MMR, rerank), entities, graph, goals, tool_memory, conversations, archivist, ingest (canonicalize→chunk→score→tree), diff, sync (composio clickup/github/linear/notion/slack/gmail), providers, persona (readers for claude_code/codex/git history), fsutil.

Host-facing extension points (~20 public traits) with the seam's current implementations in parentheses: Memory, MemoryStore, EmbeddingBackend + Embedder (SeamEmbedder), ChatProvider (SeamChatProvider), Summariser (HostSummariser), SealObserver + seal-time Embedder bridge (seal.rs), TreeJobSink (HostTreeJobSink), QueueDelegates (HostQueueDelegates), SourceReader, SelfIdentity, EntityOccurrenceIndex, SnapshotItemSource, GoalsGenerator, ConversationEventBus/ChannelEventHandler, PersonaStateStore, and the sync seams SyncStateStore / SyncEventSink / SkillDocSink / LocalDocumentSink / ExternalSourceReader / SyncPipeline / ActionExecutor / IncrementalSource (sync.rs, HostSyncAdapter).

The crate consumes tinyagents::harness::embeddings::{EmbeddingModel, format_embedding_signature} in store/vectors/embedding.rs — that is the whole tinyagents dependency (the W-EMB decision made real by tinyagents #58).

Crate CI: fmt, clippy -D warnings, build+test --all-features, doc build with warnings denied, plus a per-feature matrix (core, tokio, git-diff, sync, all). Offline by convention; Composio/live tests are wiremock-mocked or env-key-gated.

---

4. Audit: what remains host-side

4.1 Per-domain state (crate-backed file counts via use tinycortex::)

Domain	Files / LOC	Crate-backed	State
memory/	69 / 18.0k	3	Orchestration + policy + RPC layer (by design "no SQLite here"). Predominantly host: read_rpc/, schema*/, rpc_models, agent tools, sync orchestration, tree policy. Stays, minus shim cleanup.
memory_store/	54 / 17.7k	15	Sub-stores flipped to crate (chunks, content, vectors, kv, entities, trees, safety/PII per W3/W7 markers). unified/ (~5k + ~3.7k tests) is README-marked "Staging for removal" — documents/query/segments/events/profile await per-kind flips. factories.rs (config/provider selection) stays host.
memory_tree/	47 / 10.3k	16	Largest crate-backed surface; engine parts delegate. Host keeps RPC (tree/rpc.rs 1,387, retrieval/rpc.rs), CLI, bus subscriber, health/ doctor (G5, host-retained by decision), and score/embed/ providers (~1.9k — see WP-3).
memory_queue/	7 / 1.4k	2	W4 flipped: worker.rs delegates claim→dispatch→settle to tinycortex::memory::queue::run_once via HostQueueDelegates; per-kind handlers deleted. Host keeps worker pool, scheduler tick, scheduler_gate throttle, SQLite glue. Done, minus the memory::jobs re-export shim.
memory_search/	8 / 0.8k	2	W5 flipped: scoring.rs/vector/mmr.rs are near-empty shims over crate WeightProfile/retrieval::mmr. Host keeps the three agent tools. Done.
memory_tools/	9 / 1.4k	3	W7 flipped: types/store/prompt are shims over tool_memory::{types,store,render}. Host keeps capture.rs (PostTurnHook) + put/list tools. Done.
memory_archivist/	1 / 44	1	Pure W7 shim over memory::archivist. Done.
memory_conversations/	2 / 0.9k	2	W7 shim over memory::conversations; host keeps bus.rs event-bus glue (+ a tinychannels legacy-message conversion). Done.
memory_diff/	7 / 2.0k	11 uses	W7 flipped: engine crate-owned; host keeps RPC/tools + DomainEvent::MemoryDiff* publishing. Done.
memory_goals/	6 / 0.9k	4 uses	W7 flipped: types/store crate-owned (byte-identical MEMORY_GOALS.md); host keeps agent wiring. Done.
memory_sources/	16 / 5.0k	12 uses	Product registry (config-persisted connectors) with crate readers underneath. Host by design.
memory_sync/	73 / 17.7k	5 uses	The largest remaining mass. Host Composio provider pipelines (gmail/slack/github/notion/linear/clickup) still run host-side while the crate's memory::sync ships the same provider engine. W-SYNC seam (sync.rs, 679) is landed; the host flip (W-SYNC.3) has not happened. See WP-2.
embeddings/	13 / 3.1k	0	Bridges tinyagents (9 files), not tinycortex. Post-#58, the host per-provider adapters duplicate crate harness::embeddings providers. See WP-4.
agent_memory/, subconscious/ (memory parts), learning/ (memory parts)	—	0	Host product consumers of the memory stack; no crate duplication. Stay.

Consumer counts (host memory tiers are deeply embedded and stay host, so these are context, not deletion blockers): memory:: 166 files, memory_store:: 132, memory_tree:: 67, memory_queue:: 18, memory_tools:: 11, memory_search:: 3.

4.2 The seam — src/openhuman/tinycortex/ (11 files, ~3.2k LOC)

config.rs, embeddings.rs, chat.rs, summariser.rs, ingest.rs, seal.rs, sync.rs (679), queue_driver.rs (1,007), persona.rs (446), parity.rs (#[cfg(test)] format pins), mod.rs. Healthy and intentionally host-owned. Two findings:

• The seam is not the funnel. mod.rs re-exports crate types (MemoryEntry, MemoryTaint, RecallOpts, …) as the intended type-unification funnel, but zero files import crate::…::tinycortex:: re-exports — every memory domain imports tinycortex::memory::* directly. Either convention works; pick one and record it (WP-5). Direct crate imports are simpler and match how tinyagents is consumed; if so, shrink the facade instead of promoting it.
• Stale internal doc: mod.rs:28-30 names sibling adapters sinks.rs and bus.rs that were never created (their roles landed as ingest.rs/seal.rs/sync.rs). queue_driver.rs:21-26 still says "This brick is additive: nothing is flipped yet" — the flip landed in #4781.

Coupling is bidirectional by design (seam reaches back into memory_store/memory_tree/memory_sync to do host work inside crate callbacks); not a defect, but queue_driver.rs's 11 host imports will thin as WP-1/WP-3 shims retire.

4.3 Broken/dangling references (resolved in WP-0)

Resolved issue	Resolution
Missing TinyAgents companion-spec link	Replaced with the existing July 22 migration plan.
Missing standalone deletion-ledger link	Replaced with §2 of the TinyCortex migration spec.
Wrong agent-harness documentation path	Pointed at the architecture subdirectory.
Never-landed golden-workspace generator	Removed after the post-cutover parity decision descoped synthetic pre-cutover fixtures.
Wrong raw-coverage test path	Corrected to the tests/raw_coverage/ location.
Stale line anchors Cargo.toml:82, src/openhuman/mod.rs:130	tinycortex-migration-spec.md §0.4 (actual: Cargo.toml:116, mod.rs:140)
Never-created seam sibling names	Replaced with the adapters that actually landed.
Developer-specific absolute path in crate docs	Replaced with repository-relative paths.

Also stale-in-place: all five docs' status headers (pre-#4794), the drift ledger's pin anchors, and the crate migration doc's claim that "TinyCortex will not own memory sync" (superseded by the plan's §8 sync-inclusive amendment and the crate's shipped sync module — the scheduler/credentials/ RPC stay host, the engine is crate; restate the boundary in one place).

4.4 Test inventory

• In-crate (done): engine coverage was ported by #4820; the crate runs its own suite offline (wiremock Composio mock; live tests env-gated). W8's CI wiring exists: ci-lite.yml:859 runs the vendor suite on submodule-pointer changes — but with --features git-diff,sync only; persona is untested in host CI while enabled in the host build.
• Dies with WP-1: memory_store/unified/*_tests.rs (~3.7k LOC: documents_tests 1,275, query_tests 1,000, profile_tests 763, segments_tests 393, events_tests 287) go with the unified tier; port any still-unique assertions to the per-kind crate backends first.
• Retargets with WP-2: the memory_sync suites (389 inline tests across 35 files + provider *_tests.rs: github 655, gmail 301+354, slack 180, linear 172, clickup 155, notion 119) — post-flip, provider engine behavior is crate-tested; host keeps orchestration/bus/config tests. Integration: memory_sync_pipeline_e2e.rs (570), memory_golden_parity_e2e.rs (275), raw_coverage/memory_sync_* retarget to the crate-backed path.
• Golden parity harness: comparators 1 & 5 green; 2/3/4 TODO; frontmatter_parity Layer-1 asserter pending; fixture generator script missing. Decide finish-vs-descope in WP-0 (post-cutover, parity pins matter mainly for MemoryTaint/on-disk stability — keep those, consider dropping the rest).
• Stays host: json_rpc_e2e.rs memory sections, memory_sources_e2e (837), memory_tree_summarizer_e2e (583), memory_roundtrip_e2e, memory_graph_sync_e2e, memory_fast_retrieve_e2e, autocomplete_memory_e2e, memory_artifacts_e2e, the 15 raw_coverage/memory_* modules, and all subconscious/learning suites.

---

5. The plan — five work packages

Same conventions as the tinyagents companion plan: WP-0 first (hygiene, unblocks honest review), then largely independent packages; drift-ledger rows before deletions; failing-before/passing-after regression tests; small validated commits on feature branches per submodule.

WP-0 — Version/tag + CI + docs housekeeping (no engine behavior change)

1. Establish the tag discipline the Cargo comment promises: tag vendor/tinycortex (e.g. v0.1.1 at the current release point, or cut v0.2.0 at HEAD), and record the bump workflow the plan prescribed (chore(vendor): bump tinycortex … — zero such commits exist in host history). Alternatively amend the comment to describe reality (main-tracking submodule). Either is fine; the current state (comment demands tags, none exist) is the worst option.
2. Align the nested tinyagents: when the tinyagents plan's WP-0 bumps vendor/tinyagents to v2.1.0, bump vendor/tinycortex/vendor/tinyagents in lockstep and update tinycortex's tinyagents = "2" if needed; then evaluate retiring the [patch."…senamakel/tinyagents"] block and its obsolete "while tinyagents #58 is pending" comment (#58 is merged).
3. Close the persona CI hole: add persona to the vendor-suite features in .github/workflows/ci-lite.yml:859 (and confirm the crate's own feature-matrix covers persona, which today it does not list either).
4. Docs refresh: update the five docs' status headers to post-#4794/#4820 reality; update drift-ledger anchors (pin daaaf6ba+tag, note #4794/#4820/ #4863); document the persona feature + seam (currently invisible in all five docs); restate the sync ownership boundary once (engine=crate, scheduler/credentials/RPC/bus=host) and fix the crate-side migration doc's contradictory "will not own sync" line + its absolute local path.
5. Fix the §4.3 broken references, including the two code-adjacent ones: correct the seam module inventory and the queue-driver's stale pre-flip description.
6. Golden-parity decision: finish comparators 2/3/4 + frontmatter_parity
   • the fixture script, or descope to the security-relevant pins (MemoryTaint byte-identity, chunk/vector on-disk format) with the decision recorded in the parity checklist.

Exit: pins/tags/CI/docs agree with the tree; grep for the phantom paths comes back clean.

WP-1 — Retire memory_store/unified/ (the "staging for removal" tier)

1. Inventory the five remaining unified facets (documents, query, segments, events, profile — ~5k LOC) against their per-kind crate backends; flip callers facet-by-facet (the W3 sub-store pattern, already proven for chunks/content/vectors/kv/entities/trees/safety).
2. The blocker is G1 (sqlite_conn() escape hatch, ~312 refs at audit time): re-audit the current count, then either land the crate-side access the gap audit sketched or explicitly re-scope G1 to the host-retained namespace-document tables and record it.
3. Decide the fate of the 10-table host-retained UnifiedMemory namespace-document tier: keep host (spec's standing decision) but move it out of unified/ into a clearly-named home so "staging for removal" can actually be removed.
4. Tests: port unique assertions from the ~3.7k LOC of unified/*_tests.rs to per-kind backends (crate-side where engine behavior, host-side where glue), then delete with the code.

Exit: memory_store/unified/ gone or reduced to the renamed host tier; memory_store/ ≈ 9–10k LOC of glue + host tier; memory_roundtrip_e2e + memory_golden_parity_e2e + full mock-suite green.

WP-2 — W-SYNC.3: flip memory_sync/ onto the crate sync engine

The seam (sync.rs: ExternalSourceReader, SkillDocSink, LocalDocumentSink, SyncStateStore, SyncEventSink) is landed; the crate ships the Composio engine incl. per-provider modules. Remaining:

1. Close the D4.1–D4.4 drift rows (ledger requires this before the flip).
2. Flip provider-by-provider (gmail first — the crate already carries the Gmail 413-page fix #73), keeping host-side: scheduler/periodic trigger, credentials, config.toml source registry (memory_sources/), event-bus publishing (ChannelMessageReceived et al.), RPC/status surface, and redaction/source_scope gating.
3. Dedupe post-processing: host providers/*/post_process* vs crate provider modules — upstream generic normalization, keep host product policy.
4. Tests per §4.4: engine behavior moves to crate suites (wiremock Composio), host keeps orchestration/bus/e2e; memory_sync_pipeline_e2e retargets.

Exit: memory_sync/ shrinks from 17.7k toward ~6–8k LOC of orchestration/product glue; D4 rows CLOSED; no duplicated provider parsing.

WP-3 — Embedding-provider dedupe (lands in tinyagents, coordinated here)

Two host clusters now duplicate tinyagents::harness::embeddings (post-#58: openai/cohere/voyage/ollama/cloud + rate-limit/retry-after):

1. src/openhuman/embeddings/ adapters (openai/voyage/cohere/ollama/ cloud_adapter.rs, provider_trait.rs) → construct crate EmbeddingModels directly; host keeps rpc.rs (1,479), schemas.rs, catalog.rs, factory.rs (config/BYOK selection, #4056 dimension gating).
2. memory_tree/score/embed/{openai_compat,ollama,cloud,factory}.rs (~1.9k) → same crate models via the seam's SeamEmbedder/EmbedderBridge.

Because tinycortex consumes the same EmbeddingModel trait, this single change serves both crates — it is the concrete payoff of the W-EMB decision. Sequence after WP-0's tinyagents version alignment.

Exit: one embedding-provider implementation per provider across the workspace; embeddings_rpc_e2e (822) + ollama_embeddings_fallback_e2e (234) green against the crate-backed path.

WP-4 — Shim retirement + funnel decision

1. Funnel policy: adopt direct tinycortex::memory::* imports as the convention (matches reality and the tinyagents precedent), shrink the seam mod.rs facade to the types host code genuinely re-brands (MemoryTaint stays pinned by parity tests) — or the reverse; either way, record it in the spec and stop carrying both.
2. Retire compatibility shims whose consumers can move: memory_tree/tools.rs (points at memory), memory_store/trees legacy paths, the memory_queue→memory::jobs re-export, memory_archivist (44-line shim — move its 2 call sites and delete the domain), and the W5/W7 near-empty shims once import counts hit zero (memory_search/scoring.rs, vector/mmr.rs, memory_tools/{types,store}.rs).
3. Thin queue_driver.rs (1,007) as WP-1 removes its memory_store reach-backs.

Exit: shim count measurably down (each deletion = one commit with its import-migration); seam ≤ ~2.5k LOC; no module whose entire body is a re-export except deliberate facades recorded in the spec.

WP-5 — Exit gate

• Full suite: scripts/test-rust-with-mock.sh (incl. the 15 raw_coverage/memory_* modules), vendor suite cargo test --manifest-path vendor/tinycortex/Cargo.toml --features git-diff,sync,persona, slim disabled build (--no-default-features --features tokenjuice-treesitter — memory domains are ungated but the standing repo rule applies), pnpm rust:check.
• Drift ledger: D4 CLOSED; new rows for every WP-1/WP-2 deletion; the spec's §2 deletion-ledger skeleton filled in with actuals.
• Docs: the five tinycortex docs + src/openhuman/memory*/README.md files + gitbooks/developing/architecture.md reflect the post-flip reality; this plan stamped done.

Execution record (2026-07-22)

Package	Result
WP-0	TinyAgents unified at 2.1 across OpenHuman, TinyCortex, and TinyFlows; obsolete fork patch removed; TinyCortex CI covers persona; docs and phantom references repaired. TinyCortex intentionally tracks reviewed upstream commits rather than nonexistent tags.
WP-1	G1 count is zero. The ten-table namespace/document product tier is host-owned, so the complete unified/ directory was renamed to namespace_store/ instead of deleted as engine duplication. Its 153 targeted tests pass.
WP-2	The live default path was already TinyCortex-backed. D4 is CLOSED; dead Gmail duplicate removed; remaining provider task/profile projections explicitly classified as product policy. Provider tests pass (301).
WP-3	Concrete provider transports deduplicated into TinyAgents. The memory tree now uses a thin ProviderEmbedder; Ollama's 8k context/batch and missing-model guidance moved upstream before the host client was deleted. TinyAgents embedding tests pass (38); host library check passes.
WP-4	Archivist, search scoring/MMR, tool-memory type/store, tree-tool, jobs-alias, and unused seam type facades retired. Direct crate imports are canonical. Seam production code is 2,229 LOC (below the 2.5k exit target).
WP-5	Local focused validation is recorded above; the slim --no-default-features --features tokenjuice-treesitter build also passes. CI Full run 29925645209 is green: core quality and full tests, TinyCortex, Tauri, frontend, mock-backend Rust E2E, Playwright, three desktop builds, every launched desktop shard, and the final gate. Two first-attempt Linux jobs ended without uploaded logs; rerunning only failed jobs passed the Linux build, Rust integration suite, and all eight Linux shards.

---

6. Risks and standing gotchas

• MemoryTaint is security-critical. The type is proven byte-identical host↔crate and fails closed to ExternalSync; the parity pins in tinycortex/parity.rs are the regression guard — whatever WP-0 decides about the golden harness, keep these.
• Redaction and source_scope gating stay host. WP-2 must not let crate sync code become a path around host redaction or the source allowlist — review the SyncEventSink/LocalDocumentSink implementations for every provider flipped.
• On-disk compat: chunk/vector formats, MEMORY_GOALS.md byte-identity, and the P5 fresh-DB divergence (3 legacy inline-embedding columns in mem_tree_chunks) — migrations of user data must be forward-only and tested against a pre-migration workspace fixture.
• MIT vs GPL: tinycortex is MIT (tinyagents is GPL-3.0) — moving host code down into tinycortex relicenses it; fine for our own code, but keep the "no product policy / no keys / nothing openhuman-branded as API" rule, and don't move anything derived from GPL-only sources into the MIT crate.
• Two Cargo worlds + nested submodule: vendor bumps now touch up to three lockfiles (root, app/src-tauri, and tinycortex's own) and two submodule pointers (vendor/tinycortex, vendor/tinycortex/vendor/tinyagents). CI clones with submodules: recursive for the vendor suite — verify after any bump.
• Event-bus surface: memory_sync publishes 7 distinct DomainEvent variants; the WP-2 flip must keep publishing them from host callbacks (the crate has no event bus by design).

---

7. Quick reference — disposition of every audited module

Module	Verdict
memory_store/unified/	DELETE after per-kind flips; host tier re-homed (WP-1)
memory_store/ sub-stores, factories.rs	DONE (crate-backed) / STAYS (host glue)
memory_sync/ provider engines	FLIP to crate memory::sync (WP-2)
memory_sync/ scheduler, bus, RPC, credentials	STAYS
embeddings/ per-provider adapters	REPLACE with tinyagents::harness::embeddings models (WP-3)
embeddings/ rpc/schemas/catalog/factory	STAYS
memory_tree/score/embed/ providers	REPLACE via seam + crate models (WP-3)
memory_tree/ engine files	DONE (crate-backed); RPC/CLI/bus/health STAY
memory_queue/, memory_search/, memory_tools/, memory_diff/, memory_goals/, memory_conversations/	DONE — retire residual shims when import counts reach zero (WP-4)
memory_archivist/	DELETE the 44-line shim after moving 2 call sites (WP-4)
memory/ (orchestration/RPC/tools/tree-policy)	STAYS
memory_sources/	STAYS (product registry over crate readers)
agent_memory/, subconscious/, learning/	STAYS (consumers, no duplication)
Seam src/openhuman/tinycortex/	STAYS, shrinks; fix stale module docs (WP-0), thin queue_driver.rs (WP-4)
Golden-parity harness	FINISH or DESCOPE by decision (WP-0); MemoryTaint/format pins KEEP