Files
openhuman/docs/tinycortex-migration-spec.md
T

12 KiB
Raw Blame History

TinyCortex Memory Migration — Spec (Phase 0.5 / 0.6)

Status: Phase 0 baseline. Anchors the migration to exact reviewed SHAs and ties together the drift, gap, and parity ledgers. Modeled on docs/tinyagents-migration-spec.md + its deletion ledger.

Companion plan: tinycortex-memory-migration-plan.md Ledgers: tinycortex-drift-ledger.md · tinycortex-api-gap-audit.md · tinycortex-parity-checklist.md

Version anchors

Repo SHA Note
tinyhumansai/openhuman 7850cf363559bcbb7ba688cbc4fccdb6bd9ce754 host audit base (main, 2026-07-04)
tinyhumansai/tinycortex d1a8c7be2babc8fff7a72ed93861f459f3d6fa58 crate audit base (v0.1.1)
tinyhumansai/tinycortex 35f9518 (branch chore/align-native-deps-for-openhuman) first required upstream PR — native-dep alignment (§0.4); host may not activate the dep until this merges

Port line (derived by content, §0.1): after 2026-06-25, before 2026-06-28 for engine features.


0.5 — Type-unification decision

Decision: host re-exports the crate types (pub use tinycortex::memory::{…} from memory/traits.rs), the preferred option in the plan. One source of truth; 30+ consumer sites keep their import paths through the re-export. Rationale confirmed by 0.3 (wire-compatible on disk) and the MemoryTaint proof below. Fallback (host types + From conversions) is not needed — no serde/API divergence was found.

MemoryTaint — security-critical, proven identical (required before re-export)

MemoryTaint drives external-effect-tool gating (a tainted subconscious turn must refuse external_effect tools). Its serde form, db strings, and fail-closed default were compared byte-for-byte:

Property Host (memory/traits.rs:25) Crate (types.rs:26) Match
Variants Internal (#[default]), ExternalSync Internal (#[default]), ExternalSync
serde snake_case snake_case
as_db_str internal / external_sync internal / external_sync
from_db_str unknown ExternalSync (fail-closed) ExternalSync (fail-closed)

The more restrictive taint is the default and the unknown-decode target on both sides — the fail-closed-to-ExternalSync invariant is preserved. Re-exporting MemoryTaint from the crate does not weaken provenance. A dedicated seam test (W2) pins this: unknown db string → ExternalSync, Default::default()Internal, and round-trip of both db strings.

sqlite_conn() escape hatch (W2 sub-decision)

The host Memory trait's sqlite_conn() (gap G1) is not part of the re-exported crate trait. Keep it as a host-side extension trait during the transition; migrate internal raw-SQL callers to tinycortex::memory::chunks::with_connection in W3; drive the residual count to zero in the deletion ledger. Re-export covers the data types (MemoryEntry, MemoryCategory, MemoryTaint, RecallOpts, NamespaceSummary, GraphRelationRecord, RetrievalScoreBreakdown, NamespaceMemoryHit, …) and the Memory trait's async CRUD surface; the escape hatch stays host until W3 retires it.


0.4 — Toolchain baseline (result)

Check Result
Crate edition 2021 (matches host)
[patch.crates-io] tinycortex = { path = "vendor/tinycortex" } pre-staged in both worlds (root + app/src-tauri)
CI submodule checkout recursive on all build/test lanes; covers vendor/tinycortex (verify release lanes in W1, as tinyagents needed)
rusqlite alignment ⚠️ blocker resolved upstream. Crate pinned 0.32 (bundled), host pins =0.40.0 (bundled). Two links = "sqlite3" = hard Cargo error. Fixed in the crate PR (bump to 0.40 + usizei64/try_from).
git2 alignment ⚠️ blocker resolved upstream. Crate pinned 0.19, host 0.21 (vendored-libgit2). Two links = "git2". Fixed in the crate PR (bump to 0.21 + API deltas: Tag::message, StringArray::Iter, Buf::as_str).
Crate compiles with aligned deps cargo check --all-targets clean; 38 diff/checkpoint tests pass.
Host root world compiles with dep active cargo check --manifest-path Cargo.toml --lib exit 0 with tinycortex = "0.1" active + submodule at 35f9518. No multiple packages link to native library error — one bundled SQLite + one libgit2 confirmed. (Verified locally; the Cargo.toml/lock activation is reverted from the Phase-0 docs branch and re-lands in W1 post-upstream-merge.)
Host app/src-tauri world to verify in W1 (separate Cargo world / lockfile).
GGML_NATIVE=OFF macOS ARM to verify on a macOS runner in W1 (no macOS host here).

Activation is deferred to W1. Per the submodule rule, the host may only bump the gitlink to a merged upstream SHA. The native-dep alignment is committed on the crate branch and ready to PR; once it merges and is published, W1 lands: (a) chore(vendor): bump tinycortex, (b) [dependencies] tinycortex = "0.1" in both worlds. The [dependencies] line and the compile verification below were validated locally against the branch to prove the baseline is sound.


1. Ownership split (canonical, refined by the audits)

Moves to TinyCortex (delete from host after cutover)

Host module Crate counterpart Substrate tables that move
memory_store/{chunks,content(core),vectors,kv,entity_index,safety} store/, chunks/ mem_tree_chunks, mem_tree_chunk_embeddings(+reembed_skipped), vectors, kv_global, kv_namespace, store_meta, mem_tree_entity_index, mem_tree_entity_edges, mem_tree_entity_hotness, mem_tree_ingested_sources, mcp_writes, legacy_marker
memory_tree/{tree,retrieval,score,summarise} tree/, retrieval/, score/ mem_tree_trees, mem_tree_summaries(+embeddings,+reembed_skipped), mem_tree_buffers, mem_tree_score
memory_queue/ queue/ mem_tree_jobs
memory/ingest_pipeline.rs internals ingest/
memory_diff, memory_entities, memory_graph(engine), memory_goals, memory_archivist, memory_sources(registry + local readers), memory_tools(engine), memory_conversations(engine), memory_search(vector,scoring) same-named crate modules
memory/traits.rs core types tinycortex::memory::{…} (re-export)

Stays in OpenHuman (product policy, I/O, surfaces)

  • RPC surfaces: memory/{ops,schemas,schema,read_rpc}, rpc_models.rs. Method names/payloads unchanged.
  • Agent tools: memory/tools/, memory/query/, memory_search/tools/, memory_tools(tool surface) — thin wrappers over crate retrieval + SecurityPolicy gating.
  • Live sync: all of memory_sync/, memory/sync.rs lifecycle + bus stage events.
  • Process glue: memory/global.rs singleton + queue worker; memory/source_scope.rs task-locals; memory/chat.rs; embeddings provider wiring.
  • Policy/UX: preferences.rs, remember.rs, tree_policy.rs, util/redact.rs, config mapping.
  • Host-retained UnifiedMemory namespace-document tier (0.3 key finding) — the 10 tables that coexist in the shared DB but do not move: memory_docs, graph_global, graph_namespace, episodic_log (+ episodic_fts + triggers), event_log (+ event_fts, event_embeddings, triggers), conversation_segments, segment_embeddings, vector_chunks, user_profile. These live in memory_store/unified/{init,fts5,events,segments,profile}.rs and remain host — the crate is the primitive substrate, not a drop-in for the whole DB.
  • Content-store host surfaces the crate explicitly excludes: content::wiki_git, content::obsidian, content::obsidian_registry.

The adapter seam: src/openhuman/tinycortex/ (W1, mirrors src/openhuman/tinyagents/)

embeddings.rs (EmbeddingBackend/Embedder), chat.rs (ChatProvider/Summariser×2/ EntityExtractor/GoalsGenerator), queue_driver.rs (QueueDelegates + tokio worker loop + Sentry/bus), config.rs (ConfigMemoryConfig), sinks.rs (TreeJobSink/TreeLeafSink/ SnapshotItemSource/EntityOccurrenceIndex), bus.rs (engine outcomes → DomainEvent), mod.rs (facade re-exports + boundary doc). All 17 seam traits confirmed present (§0.2).


2. Deletion ledger (skeleton)

Every legacy engine file is deleted only when its module's drift row is closed, its gaps are resolved, and the golden-workspace parity harness is green for its flip. Counts from the host audit SHA.

Legacy module Files (test files) Deletes in Preconditions
memory_store/ 66 (11) W3 drift D3 closed; gap G1 (escape hatch) migrated to with_connection; parity P3/P5/P11/P12 green; unified/ tier re-homed as host-retained (kept, not deleted)
memory_tree/ 65 (7) W5 gaps G3 (seal-embed), G6 (2× Summariser) resolved; source_scope allowlist re-verified; parity P7/P11 green; health/ + tree_policy.rs kept host (G5)
memory_queue/ 10 (1) W4 drift D2 closed (predicate upstreamed); job payload_json parity (P4/P9); host worker loop + Sentry/degraded wiring kept host
memory_conversations/ 7 (1) W7 drift D1 closed; bus.rs kept host
memory_diff/ 7 (0) W7 git-ledger parity (P9) green
memory_entities/ 3 (0) W7 parity P8 green
memory_graph/ 3 (0) W7 gap G2 resolved (derive-on-read parity vs host-retained graph_*)
memory_goals/ 7 (0) W7 seam GoalsGenerator wired
memory_archivist/ 6 (0) W7 TreeLeafSink seam wired
memory_sources/ 16 (0) W7 registry + local readers move; live sync stays host
memory_tools/ 10 (1) W7 engine → tool_memory/; tool surface kept host
memory_search/ 8 (0) W5 vector/scoring → crate retrieval/score; tools/ kept host
memory/ingest_pipeline.rs internals (thin entry points kept) W6 ingest_chat/ingest_document_with_scope signatures unchanged; 11 call sites untouched

Kept host (never deleted): memory/{ops,schemas,schema,read_rpc,tools,query,tree_source, ingestion,util}, memory/{global,source_scope,chat,sync,preferences,remember,tree_policy,rpc_models, traits(→re-exports)}.rs, all of memory_sync/, memory_store/unified/* (the namespace-document tier), memory_store/content/{wiki_git,obsidian,obsidian_registry}, memory_tree/health/, and the new src/openhuman/tinycortex/ seam.

3. Workstream order (one workstream ≈ one host PR)

W1 seam scaffolding → W2 types/trait re-export → W3 store+chunks → W4 queue → W5 tree+retrieval+score → W6 ingest → W7 long tail → W8 test-port + golden parity sweep + deletion-ledger close-out.

Each risky workstream is a sandwich (plan §4): (a) tinycortex PR(s) closing that module's drift/gap ledger, (b) host chore(vendor): bump tinycortex, (c) host cutover PR (adapter flip + legacy deletion + host-side tests in the same PR for the ≥80% diff-coverage gate).

4. Security review items (must have dedicated seam tests)

  1. MemoryTaint fail-closed to ExternalSync — proven identical (0.5); pin with a W2 seam test.
  2. source_scope per-turn allowlist — must survive the W5 retrieval cutover; the retrieval primitives (query_source/query_topic/drill_down) run inside the host's task-local scope, and a W5 seam test must assert an out-of-allowlist source is not returned.