mirror of
https://github.com/mastra-ai/mastra.git
synced 2026-11-17 08:44:46 +00:00
2 commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
|
b40fa91496 |
refactor(memory): curate observations directly in Subconscious (#22738)
## Summary Simplify experimental Subconscious knowledge ingestion around two built-ins, `remind` and `curate`, both configured in the observation agent list: - keep `remind` unchanged for selective retrieval of relevant durable knowledge; - make `curate` the sole experimental Subconscious ingestion and reconciliation agent, receiving each successfully committed observation directly; - remove the separate `capture` and `learn` agents, reflection-time Subconscious execution, and Subconscious-owned worklist/cursor/cadence orchestration. This is intentionally deletion-heavy to reduce two ingestion prompts and quality surfaces to one: 40 files changed, with 1,047 additions and 4,032 deletions overall; Subconscious production source has 80 additions and 718 deletions. This is plan-driven work with no linked GitHub issue. ## How direct curation works Observation strategies preserve the raw per-cycle observation delta separately from cumulative observational memory. After persistence and observation end markers succeed, the shared strategy lifecycle awaits one failure-isolated curator invocation per nonblank source-thread delta before normal reflection, so curation model latency is now part of every observation-cycle path that produces a delta. Resource-scoped cycles retain per-thread attribution and additionally run one curator call per contributing thread delta, serially. The curator receives the current observation delta, searches relevant existing knowledge through constrained read tools, and creates or reconciles durable knowledge through code-governed write tools. Provenance, timestamps, source identity, scope ceilings, versioning, activity, mention indexing, and semantic indexing remain code-owned rather than model-supplied. Reminder execution remains independent: reminder failure or no-signal cannot suppress curator delivery. Curator failure is reported and isolated after observation commit; it does not fail the committed observation or prevent normal reflection. ## Deliberate reliability boundaries Direct curator delivery is best-effort in this PR. Removing the durable worklist/cursor frontier means a failed or interrupted curator run is reported but is not automatically replayed. This PR does not introduce a replacement queue, scheduler, retry, cadence, or backoff mechanism. Core storage worklist/cursor APIs, including `knowledgeBySource`, remain for other consumers such as Factory's knowledge route; the Subconscious runtime stops using them for curation orchestration. The curator continues to require `organizationId` in request context, matching the prior capture path; a missing value is reported as a curator error after observation commit. The new `knowledge_create` tool preserves the existing node-first storage sequence: if the initial record append fails after node creation, the node may remain. No new cross-adapter transaction contract is introduced. ## Compatibility and cross-package cleanup - `@mastra/memory`: removes `capture`, `learn`, reflection-time Subconscious execution, worklist/cursor/cadence orchestration, and the public `Memory.runCuration()` method. There is no replacement manual curation entry point; Subconscious curation now runs only from successfully committed observations. - `@mastra/code-sdk`: removes Factory-only capture-pinning and cadence configuration. - `@mastra/factory`: removes phase-exit curation triggers that depended on `Memory.runCuration()`. - Simulator: removes capture arms, cadence/tail flushing, cursor accounting, and A/B comparison in favor of one direct-curator replay path. A minimal patch changeset is included for `@mastra/memory`, `@mastra/code-sdk`, and `@mastra/factory`. ## Verification - Focused Subconscious, lifecycle, replay, reminder, scope, activity, read-tool, and knowledge-write coverage: 11 files, 59 tests passed. - Full `@mastra/memory` unit suite: 63 files, 1,497 tests passed, 1 todo. - Memory simulator suite: 3 files, 46 tests passed. - `pnpm build:memory`: 15/15 tasks passed. - Memory typecheck and lint passed; lint reported 0 warnings and 0 errors. - `@mastra/factory` build, typecheck, full tests, and lint passed: 103 files, 1,980 tests passed, 6 skipped. - `@mastra/code-sdk` build, typecheck, and lint passed; the changed memory-construction suite passed 15/15 tests. - Red/green lifecycle coverage and a deterministic two-cycle replay over real in-memory storage verify direct delivery, reconciliation to one active fact, later cross-thread reminder retrieval, and no Subconscious worklist/cursor calls in the exercised paths. No live hosted-model run is claimed. - Removal greps, `git diff --check`, and clean-status checks passed. Two broader gates could not complete cleanly; details follow. ## Verification limitations The isolated `packages/memory/integration-tests` bootstrap could not install dependencies in this environment: the frozen install reports an override mismatch, and the required non-frozen retry cannot fetch private `@internal/ai-v6` from the npm registry. The integration tests therefore did not execute. The full Code SDK suite completed with 105 files and 1,524 tests passing, while four tests failed from startup timeout, `spawn E2BIG`, callback-port `EADDRINUSE`, and an MCP OAuth callback assertion. These failures were not used as evidence for this change; the changed SDK memory-construction suite passes independently. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## ELI5 The PR makes `curate` the single agent that turns new observations into knowledge. It removes older capture, learn, and automatic scheduling paths so curation runs independently after observations are saved. ## Changes - Refactor Subconscious ingestion around `remind` and `curate`. - Dispatch committed observation deltas directly to `SubconsciousCurateExtractor`. - Make curation asynchronous and failure-isolated. - Remove reflection-time execution, capture and learn agents, curation cadence, worklist orchestration, and cursor management. - Preserve reminder execution, provenance, scope enforcement, and code-owned indexing. - Add `knowledge_create` with scoped writes and provenance metadata. - Validate RFC 3339 timestamps for knowledge writes. - Harden observation framing against untrusted input. - Remove related public APIs and types from Memory, Code SDK, Factory, and simulator integrations. - Replace the simulator A/B workflow with direct observation replay. - Add tests for replay isolation, curation behavior, scope enforcement, provenance, failure handling, and input validation. - Add a patch changeset for `@mastra/memory`, `@mastra/code-sdk`, and `@mastra/factory`. ## Validation Focused tests, memory and simulator suites, builds, typechecks, linting, and relevant Factory and Code SDK checks passed. Memory integration tests were blocked by dependency installation issues. Four unrelated failures remained in the full Code SDK suite; changed SDK tests passed independently. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: Jamie Nichols <CodingButter@users.noreply.github.com> Co-authored-by: Mastra Code (openai/gpt-5.6-sol) <noreply@mastra.ai> Co-authored-by: Jamie Nichols <10267569+CodingButter@users.noreply.github.com> Co-authored-by: Tyler Barnes <tylerdbarnes@gmail.com> |
||
|
|
eda942cb76 |
feat(memory): add an observational memory replay simulator (#22058)
## Summary Add local conversation simulation tooling to `@mastra/memory` for comparing experimental Subconscious capture and curation prompt changes against reconstructed real observation cycles. ## Root cause Evaluating prompt changes through deployments mixes prompt behavior with newly derived cycle boundaries and requires days of organic usage. Re-observing raw messages would also let the tested prompt move segmentation boundaries, making A/B results look meaningful while changing more than one variable. The simulator tests are colocated with the implementation they cover under `packages/memory/scripts/simulate`, which sits outside the package's original `src/**/*.test.ts` Vitest include. The package include is widened to `['src/**/*.test.ts', 'scripts/**/*.test.ts']` so they are discovered and run as part of the normal unit suite. ## What changed - Add local extract, replay, and A/B commands under `packages/memory/scripts/simulate`. - Reconstruct persisted observation cycles while excluding reflection heads from later generations. - Reject duplicate `generationCount` records as ambiguous instead of replaying them in unstable order. - Preserve deterministic ordering and report missing-generation warnings. - Replay through the real capture and curation path with optional production knowledge-resource scoping. - Isolate every arm in a fresh local database and refuse remote write targets. - Add an A/A control to expose model sampling noise. - Keep the simulator suites beside the implementation under `packages/memory/scripts/simulate`, named after the module each covers (`diff.test.ts`, `extract.test.ts`, `replay.test.ts`, `reconstruct.test.ts`, `drive.test.ts`, `drive.isolation.test.ts`), and widen the package Vitest include so they are still collected by the unit run. - Add simulator TypeScript checking to package verification. - Export the existing observation-boundary regex for simulator reuse without changing its pattern or runtime behavior. ## Design notes This is a bounded simulator, not a general PostgreSQL cloning tool. Generated target DDL does not preserve every PostgreSQL type, constraint, default, or precision detail. Worklist counting is capped, missing generations produce incomplete comparisons with warnings, and the content differ remains aggregate and positional in several cases. The replay command dynamically imports the built PostgreSQL adapter, so linked workspace packages must be built before running the simulator from a fresh checkout. The README documents the local PostgreSQL, pgvector, credentials, and build prerequisites. The widened Vitest include makes the simulator suites part of the default `@mastra/memory` unit run while leaving them colocated with the tooling. Normal production TypeScript scope remains unchanged because package TypeScript excludes `**/*.test.ts`; simulator checking remains owned by `tsconfig.scripts.json`. ## Safety - Source access is read-only. - Replay targets must use a literal IPv4 or IPv6 loopback address; hostnames and routing query parameters such as `host`, `hostaddr`, and `service` are rejected. - Extraction, standalone replay, and database recreation attest the live PostgreSQL server address as loopback before target initialization, writes, or destructive operations. - Extraction compares connected database identity and port, treating IPv4 and IPv6 loopback as the same local server, and refuses to overwrite the source database. - Each arm uses a separate fresh database. - No conversation or knowledge payload is committed. Tracking: COR-1167 ## Test plan From `packages/memory`: ```bash pnpm vitest run scripts/simulate --reporter=dot pnpm test:unit pnpm lint git diff --check ``` Results on the reviewed local head: - Simulator selection: 78 passed across 6 files. - Full memory unit suite: 1,602 passed and 1 todo across 68 files. - Lint: 0 warnings and 0 errors across 207 files. - Diff check: passed. - Simulator suites are discovered at their colocated path by the widened include; the full unit run collects 69 files with the suites in place. The production-source blast radius is limited to exporting the existing observation-boundary regex. Its pattern and runtime behavior are unchanged. No changeset is included: this adds repo-local tooling under `scripts/`, not a shipped package feature. --------- Co-authored-by: Jamie Nichols <CodingButter@users.noreply.github.com> Co-authored-by: Jamie Nichols <10267569+CodingButter@users.noreply.github.com> Co-authored-by: Mastra Code (openai/gpt-5.6-sol) <noreply@mastra.ai> |