Changelog

Python SDK 2.7.0 and TypeScript SDK 2.6.0: search and reason request the compact view by default (pass view "full" or debug for diagnostics), search adds prefer_entity_match and include_profile, and chat accepts answer_style and no_information_reply.

/v1/chat/ask uses a stored condition, restriction or plan that bears on the question (for example a mobility need when asked for weekend ideas) instead of declining because that fact does not answer the question directly.

The first requests after a deploy or a worker restart no longer wait for language and keyword-index models to load: workers load them when they start.

Chat answers personal questions in plain language from the stored facts by default. answer_style "quote" returns the stored text word for word instead, and no_information_reply sets the reply used when nothing stored answers the question (default "I don't know that yet."). A stored condition, restriction, preference or plan that bears on a request is now used even when it does not answer it directly; it appears in memory_context.memories with the grounding reason relevant_user_context, and memory_context.answer_composition describes how the answer was put together. degraded_stages can report user_context_judgement_unavailable and empty_model_reply. The Python SDK's MemoryChat.send() takes both fields, and the TypeScript SDK adds chatAnswerFields() for OpenAI-compatible clients.

Chat learning reports when learned facts become searchable: memory_learning is returned at the top level with readiness, searchable, estimated_ready_seconds, estimated_ready_at and estimate_basis. Learned facts are usually searchable within 4 to 5 seconds.

Search sets low_confidence and abstain_recommended only for genuinely weak or unverified matches in the default soft grounding, and grounding.status is "not_applicable" when there is no claim to verify. Every result carries applies_to and entity_match (match, partial, mismatch or unscoped) against the response's query_entities, and prefer_entity_match ranks mismatches last. include_profile returns the user's standing facts and preferences in a separate profile field. view "compact" is slimmer, debug restores the diagnostics, and both SDKs now default to compact.

Every extracted memory has a memory_type (rule, procedure, preference, fact or event), returned by GET /v1/memories, by process job events and in search metadata, and search can filter on it with memory_types. Agent transcripts are stored as general rules plus one event per task, and a process job that finds no fact reports memories_created 0 instead of storing the raw message.

Reason sets decision_abstained whenever answer is null and explains why in abstain_reason (for example rule_does_not_cover_case or authorization_not_verifiable). A question on the other side of a stored threshold rule returns the rule with the computed comparison in metadata.rule_application, such as $100 is not above $250.

The public OpenAPI contract at api.hebbrix.com/openapi.json now lists only operations API customers can call (339 paths, down from 362). Hebbrix staff tools (RL training and checkpoints, legacy policy administration, system stats, support ticket administration, the graph connection test), the RL canary routes, which always return 501 Not Implemented, and the Stripe and Tawk.to webhook receivers are left out, along with the webhook signature schemes. They keep working unchanged; no customer request changes. Each route declares its audience in code, so new staff routes stay out automatically.

New guides walk through the two common integrations end to end: Chatbot memory (save each conversation, search before answering, answer with memory) and Agent memory (rules versus events, lessons from tool errors, checking rules before acting, learning which actions work). The Outcome Memory guide now opens with a plain-language overview and a short simulation you can run to watch it learn.

hebbrix.com/llms.txt and hebbrix.com/llms-full.txt describe the documentation for language models. Both are generated from the published pages and the public OpenAPI contract each time the site is built, so they list the current endpoints, parameters and SDK versions.

The docs now name the current published SDKs everywhere: Python 2.6.3 on PyPI and TypeScript 2.5.3 on npm. The versions come from the API's release manifest, so they no longer drift between pages. The latency guide adds measured production numbers and how they were taken.

Release-by-release notes that used to sit inside the docs pages now live in this changelog, described by date and capability.

Default unsupported personal recall can abstain before generation instead of guessing a job or other personal fact. Supplied introductions remain usable. The finite source-absence guard does not certify all generated answers or replace caller-owned tool and structured-output contracts.

Synchronous /memories/process accepts explicit wait_for_index with a bounded index_wait_seconds observation. It checks the original committed outbox, scope, member IDs, full-content hashes and current lifecycle; shortened public event text is not the source identity. Timeout, supersession or unavailable legacy proof never becomes searchable success or a failed source write. The same key can observe the original receipt without repeating extraction.

Use native auto selection when the server should own exploration. The LLM advisor path can disregard suggestions and is not the recommended learning selector. New helper native options are source-only; published Python 2.6.2, TypeScript 2.5.2 and MCP 0.5.14 are unchanged.

This is an engineering correction, not a new performance evaluation or proof of superiority over Thompson sampling. Historical failures remain unchanged; benchmarks, evaluations and raw validation receipts remain private.

Memory processing acknowledges the exact committed source and indexing event. A durable pending receipt is a successful write, not a searchable completion. Same-key replay retains the original result; follow the status URL before re-reading an update.

Explicit chat waits use the existing durable job, execution claim, chronological ordering and indexing barrier. A timeout reports actual pending status, not completion or a latency guarantee.

Free-text confidence exposes separate action evidence and descriptive limitations. Unknown or pending attempts are not successes. ACT remains advice and never authorizes execution.

This engineering release does not establish improved learning performance or universal model reliability. Published SDK versions remain Python 2.6.2 and TypeScript 2.5.2; additional source-only helpers require a separately published package. Benchmark material remains private.

Personal collection reads require an explicit end-user scope. Memory-off chat preserves caller instructions without hidden profile injection; requested and served model identities are disclosed, with optional strict alias admission.

Chat greetings get an ordinary model reply instead of canned text, deeper analytics become opt-in, and chat responses add retrieval_trace with the memory IDs that were considered, used in the prompt or dropped. Search stops padding results with unsupported exact-zero scores while keeping low nonzero scores and supported structural matches.

Outcome advice distinguishes a normal miss from losing the recent gate: the latest failure still demotes immediately, but recovery probation (eight consecutive successes) begins only after a previously qualified recent gate fails or an epoch requires requalification.

Job receipts distinguish queued, running, terminal and timed-out waits. Linked outcome context and compact response controls retain safety fields; ranking scores and descriptive reversal warnings are not correctness certificates.

Unambiguous English day references use a supplied source date, never worker time. Residence and employer profile admission requires the asserted property, not a merely mentioned city or organization. Unsupported forms conservatively abstain; arbitrary semantic accuracy is not established.

New SDK convenience fields are source-only unless an installed package explicitly includes them. Check the live release endpoint for runtime and published companion identities. Earlier failed checks remain failed; no benchmark/evaluation material is published.

This release adds compact evidence views, bounded chat-learning waits, owned batch receipt reads, explicit provisional-outcome confirmation and native selection controls. A timed-out wait is not completion; outcome capture remains provisional, not execution verification or selection-training evidence.

Action-specific evidence is separated from pooled confidence. Relevant counterexamples remain visible, weak personal candidates require support or explicit opt-in, and raw ranking scores are not correctness probabilities. Tool ACT remains advice, never permission or an executed action.

Stable Python 2.6.2 and TypeScript 2.5.2 expose matching integration helpers; they need this release or later on the server, and older installed packages do not gain them automatically. Pin versions and review the compatibility policy in the reference guide. The live release endpoint, not this page, identifies the current runtime.

Engineering checks do not establish model reliability, learning superiority, calibrated safety or latency guarantees. Performance objectives remain open; benchmarks and evaluation material stay private.

Follow-up responses expose the persisted observation value, polarity and stored source hash/span on first capture and replay. Missing historical witnesses remain null. Capture stays implicit, confidence 0.35 and nonfinal; it does not confirm execution, train serving selection or earn autonomy.

Advice cards identify legacy v2 versus v3 finite-horizon behavior, contextual/proposal modes and whether the declared remaining horizon was applied. Examples show a context-local decreasing horizon and explicit named tool proposal contracts. Compatibility defaults and permission guards are unchanged.

Absent component objectives are omitted from default configuration projections for older-reader compatibility. Only explicitly configured component policies carry the new field; no existing outcome or guard is rewritten.

Separately opt-in v3 success_cost_v1 samples binary success evidence and subtracts admitted actual-cost estimates, with action-bound caller cost assumptions for cold start. Exact scope and canonical final success/cost contracts are required; cost uncertainty is not modeled. Legacy composite selection remains available. Performance and monetary-return objectives remain unqualified.

SDKs remain Python 2.6.1 and TypeScript 2.5.1; MCP 0.5.14 and the populated database schema are unchanged. No benchmark or evaluation material is published. The live /v1/release endpoint is authoritative for the deployed build.

Follow-up capture namespaces its generated keys by bound decision and distinguishes scope/conflict errors. Bounded meaning extraction requires an exact source witness; witness authenticity is not semantic truth. Capture remains provisional and cannot train serving selection.

Explicit v3 adaptive_optional competes frozen observed-action predictions; policy_mixture compares finite native/proposal distributions using actual-action loss receipts. Advice requires a declared remaining horizon for a nonzero bounded pilot. Legacy defaults and ACT permission boundaries are preserved.

Python 2.6.1 and TypeScript 2.5.1 are the published patch releases. The deployment's live smoke stopped at 5 of 6 emitted checks after a checker incorrectly required a null descriptive aggregate. That failed receipt was not regraded and remaining live checks were unexecuted. Performance objectives and broad semantic accuracy remained open.

Explicit outcome-policy-v3 enables experimental outcome-derived change response and bounded declared optional-feature sharing. Old policies and unversioned setup retain v2 behavior. Full outcome history and the ACT evidence gate are not reset; required context and owner/action/metric boundaries remain mandatory.

Advice suggests a challenger rather than the observed incumbent, with an under-sampling rationale. V3 transient proposals are attenuated by observed outcomes only when the proposed action was actually selected. This is not calibrated proposer accuracy or a guarantee that a model follows advice.

Atomic value_objective setup creates required success and actual cost metrics. Exact-scope conversation follow-ups can create explicitly implicit, lower-confidence provisional evidence only with learning consent, a prior decision binding and learning-write permission. They do not confirm execution, train serving selection or earn autonomy.

Stable Python 2.6.0 bounds HTTPX below 1; TypeScript 2.5.0 ships typed setup options. Latency documentation distinguishes engineering objectives and large-history measurement methodology from measured guarantees. The live /v1/release endpoint is authoritative.

Performance objectives remain open. No benchmark is published, and no best-in-market, universal safety, generalization or learning-superiority claim is made. MCP and the populated database schema are unchanged.

Invalid opt-in tool selection contracts fail before generation. Per-choice selection and completion receipts disclose required/named misses, provider length/filter termination and unknown termination. Contract-violating or incomplete proposals are withheld even in annotation mode. Auto still permits no call; no external action is executed or authorized.

Supported fallback proposals are preserved, while unsupported explicit caller requirements cannot silently become automatic selection. Missing personal memory no longer makes arguments explicitly supplied by the caller require redundant confirmation; unstated arguments and permission are never inferred.

Python 2.6.0rc5 and TypeScript 2.5.0-rc.4 snapshot original advisor inputs before one callback and validate complete finite logging distributions without normalization, retries or fabricated outcomes. Caller probabilities remain unauthenticated; npm stable latest remains unchanged. MCP 0.5.14 is unchanged.

These are engineering guards, not qualification of automatic-proposal reliability, hybrid performance, generalization or learning superiority. Earlier failed studies remain failed. No benchmark is published; the live /v1/release endpoint is authoritative.

Opt-in tool_advice replies become status-only before both public serialization and stored history. A tool proposal, a held call and a missing call cannot be presented as an action executed by Hebbrix. Annotation preserves actual proposals; hold withholds every call, even ACT advice. This mode deliberately replaces unverified free-form prose, and never grants permission.

Explicit bounded incident measurements such as error rate is 37% no longer become action parameters. Operational amounts and resource counts remain constrained. Missing environment scope stays ASK and exposes needs_scope:[environment] without inventing a default. These are finite lexical guards, not universal semantic safety.

Automatic tool selection remains model-dependent. Use an explicit named or required tool_choice when your application requires a proposal, then independently authorize execution. The learning guide now prefers the already-shipped advisor-read/evidence/observe helper; transient priors remain experimental and unqualified for their performance targets.

No selector algorithm, threshold, historical outcome, SDK/MCP runtime or database migration changes. No benchmark, learning-superiority or general drift-reliability claim is introduced. The live /v1/release endpoint is authoritative.

Action advice binds recognized environment aliases, broader action scopes and typed tool arguments to the configured action. Operational quantities remain constrained while descriptive error codes do not become action parameters. These are finite identity guards, not universal semantic safety or execution permission.

Empty-context personal-fact examples were removed. Bounded owner-scoped related-query relevance counts are disclosed separately from execution outcomes, and narrowly admitted recommendation contrasts quote their source rather than asking the model to reinterpret votes. Relatedness is not verified equivalence; broad model compliance and learning-performance objectives remain unqualified.

An earlier reliability correction fixed document-provider routing, public SQL-error redaction, concurrent bulk retirement and a shared per-item profile retry budget. Required-context/action/metric admission precedes bounded learning-history windows. This does not establish the cause of earlier learning-performance misses.

The live /v1/release endpoint is authoritative. No benchmark, general learning-superiority or universal drift-reliability claim is introduced.

This release adds atomic policy setup, scoped descriptive reports and a Learning dashboard. Reports distinguish observed leaders from qualified improvements; unsupported causal uplift remains unavailable.

New low-risk opt-in configurations support experimental v2 defaults, transient advisor proposals and exact-context decision aging. Existing versioned configurations are not silently migrated. Performance targets and reliable model compliance with feedback have not been established.

Explicit policy/action/context identity supports bounded paraphrases while target and operation conflicts fail closed. Query-specific relevance feedback is presented separately from execution outcomes and can still be misread by a model.

Tool advice binds targets and exact JSON arguments. Default hold withholds every call, including ACT, for independent execution approval. Annotation is advisory; no tool is executed and no permission is granted.

Python 2.6.0rc4, TypeScript 2.5.0-rc.3 on npm next, and MCP 0.5.14 provide matching helpers. No benchmark is published. The live /v1/release endpoint is authoritative for the deployed build and artifacts.

This release introduced explicitly low-risk posterior sampling, owner-declared bounded priors and optional-field pooling, chronological windows, half-life settings and descriptive change diagnostics. Receipts record the finite sampler's actual behavior distribution.

Per-action evidence is visible before selection eligibility. Configured canonical action and target identities connect qualified outcomes to ASK, REVIEW and ACT advice, with independent rule checks. ACT remains advice, never execution authority.

Scoped outcome evidence now participates in final search ranking and chat context. SDK and MCP evidence-card helpers and operational guides are available. Rate-limit responses use JSON 429 with Retry-After; no production burst test or latency guarantee is implied.

This is an engineering release, not proof of learning superiority, calibrated confidence or universal safety. The live /v1/release endpoint remains authoritative. See the Outcome Memory migration guide for current defaults and compatibility.

A follow-up correction preserves scoped summary quotations through the public response envelope without certifying their truth. The initial release's live acceptance remains 43/45: one real summary-transport defect and one checker using an unsupported confidence scope parameter. Those results were not rescored.

This release separates procedure execution from recorded feedback, computes observed success rates from feedback, rejects conflicting final outcomes unless explicitly corrected, and isolates bounded confidence history by action relevance. Legacy procedure rows without feedback provenance are not relabeled as observed outcomes.

Caller-selected decisions can supply a complete behavior_probabilities vector for retrospective evaluation. Numeric evaluation readiness is separate from recommendation readiness; caller-reported outcomes and propensities are not verified randomization or permission to promote a policy.

Optional max_violation_rate constraints use a separate conservative aggregate gate; a raw per-decision guardrail threshold alone remains a reward penalty. Sparse, immature or missing evidence causes abstention. Independence and stationarity are assumptions, not runtime guarantees or safety certification.

Index-wait ingestion no longer silently queues asynchronous extraction, broad user summaries quote bounded current scoped sources, and learning decision rate limits use the existing tier configuration. Search scores remain ranking signals, not calibrated correctness probabilities.

Documentation now provides a five-endpoint outcome workflow, context-schema and normalization rules, readiness distinctions, and a generated endpoint/field-schema browser. No new bandit algorithm, cross-customer prior, batch outcome endpoint or world-leading performance claim is introduced. The live /v1/release endpoint is authoritative.

This release adds independently reviewed, source-bound fact revisions beyond the earlier typed attribute rules. The write transaction checks the exact predecessor and replacement; edits preserve unaffected clauses, concurrent activities and other subjects. Distinct newly extracted assertions survive retirement of the old fact. Semantic review remains fallible, not a guarantee that every new value replaces an old one.

Chat learning waits for earlier non-terminal jobs in the same account, collection, end-user and agent scope. Profile writes reject retired source memories and older conflicting source events. These guards address completion-order hazards; the customer's historical Austin/Denver occurrence has not been independently traced.

Reason can return independently reviewed, verbatim cited context when typed question parsing cannot establish coverage. Safety and permission questions retain decision_abstained even when useful context is returned: recalled facts do not authorize medication, food safety or deployment. Shared sentence boundaries preserve abbreviations and decimals.

Non-debug search responses omit duplicated plaintext proof manifests and repeated claim text. include_metadata:false removes per-result metadata without altering signed proof tokens, scope or lifecycle checks. Scores remain ranking signals. The live /v1/release endpoint is authoritative for deployment identity.

The preceding deployment passed 94 of 98 live checks. It exposed a lost new occupation, medical-context abstention, an allergy facet satisfied by a different dietary condition, and hypothetical evidence promoted by fallback review. That failed receipt is unchanged. The successor retains new assertions, distinguishes narrow evidence attributes, rejects hypothetical records independently, and binds review decisions to individual supplied source IDs. These are disclosed development regressions, not unseen customer holdouts.

This release aligns first-person questions with the same scoped subject used in the verifier's evidence text. Evidence thresholds are unchanged; hypothetical and different-person assertions remain excluded.

Scorer-only normalization preserves pronouns' relationships and question qualifiers, while converting broad requests into a consistent question form. It does not change tenant scope, certify an answer, or promise a universal relevance-score floor.

Reason retains separately stored deployment restrictions alongside a permitted time window, including approval requirements. Stored policy descriptions never grant execution authority.

The preceding v43 development acceptance passed 101 of 106 checks but exposed two Reason abstentions, an omitted approval constraint and a music fact ranked fourth in both modes. That failed receipt is preserved, not regraded. The live /v1/release endpoint is authoritative for the current build.

This release preserves the bounded ranked candidate pool in soft search even when one narrower claim is supported. A related allergy or a person's own name must not erase intolerance or family-member candidates. Permission questions are not certified by possession facts.

Reason transports fresh, query-and-content-bound facet evidence into its evidence ledger for direct subject–verb–object assertions. Compound chat retains candidates for each question, without treating one supported facet as proof of the entire answer. Missing, hypothetical and wrong-subject evidence still require abstention.

Explicit source statements that an activity ended can retire the matching atomic current fact, while keeping concurrent activities and historical or relative-subject facts. New occupations alone do not erase other jobs. Search candidates are scored on the same fresh ranking scale; scores are not calibrated correctness probabilities.

Corrected the preceding changelog entry to describe the final build of this release. The live /v1/release endpoint is authoritative for deployment identity. Product regression checks do not establish universal paraphrase coverage, medical safety, or learning superiority.

The final production build superseded the initial rollout with a correction to fresh verification-receipt transport. The earlier changelog named the initial rollout instead of this final build; that mismatch is corrected here.

Soft grounding preserves ranked candidates when a question cannot be verified, with explicit low-confidence, retrieval-only evidence. Strict grounding still requires support; a retrieved allergy memory is not a food-safety certification. The customer's second retest found all 26 previously reported questions and 27 of 32 new holdout questions, but exposed wrong-subset narrowing, Reason abstentions and occupation-update failures. Those observations are not a claim of universal recall.

Corrected direct first-person subject binding, owned-breed pet-name evidence, first-person reasoning evidence coverage, role-preserving message extraction, and chat current-state selection and reranking. These changes address failure mechanisms rather than promising universal paraphrase correctness.

Single-decision deletion accepts an owning API key with decisions:delete or wildcard scope; read-only keys remain denied and cross-owner resources remain inaccessible. See the privacy-controls documentation for credential requirements.

Added public DPA request instructions and a subprocessor information page based on the existing Privacy Policy. These pages are not a signed DPA, an audited vendor inventory, or a compliance attestation. Use the live release endpoint to confirm deployment identity.

This release added chat learning-consent guards, atomic extraction and updates, pollable jobs, search metadata controls and opt-in outcome evidence. Database schema and fleet capacity were unchanged.

The fixed release roster passed 89 production checks. A subsequent customer retest found 17 issues fixed and three public claims corrected, but default search retrieved the expected fact for only 12 of 26 new first-person phrasings. That limitation reopens retrieval and chat correctness; the original checks do not establish general paraphrase coverage.

Unsupported benchmark and compliance claims were withdrawn. The customer's exact SDK environment and the historical rare chat-hang cause were not established by this release.

This release introduced versioned context contracts and evidence boundaries for outcome decisions, with an additive database migration.

Release verification recorded 185 live context checks and 35 registered-verifier checks. These are behavioral checks, not evidence of learning superiority or long-term return on investment.

Public SDK stable versions are Python 2.5.0 and TypeScript 2.4.0. Python 2.6.0rc1 and TypeScript 2.5.0-rc.1 are explicit prereleases; source-only conveniences are not automatically published package features.

Hosted CI did not run because the account budget blocked it; this entry does not claim a green hosted CI run. Use the live release endpoint for the current source and deployment identity.

TypeScript single-memory writes with wait_for_index enabled now perform exactly one mutation and then poll the read-only status resource until the durable receipt is searchable and processing is complete. Abort, timeout, and terminal indexing failures preserve the original receipt and never repeat the write.

Python synchronous and asynchronous readiness deadlines now raise one typed IndexingTimeoutError across single writes, batches, updates, jobs, and inference. The exception retains durable memory, request, event, outbox, idempotency, and retry metadata so callers can resume observation safely.

The production OpenAPI endpoint now serves the exact checked-in artifact bytes with identity encoding and validates its length and SHA-256 at startup. The release endpoint publishes the same byte contract: f18a3290608ff17a5d1a2a9f3592ea4a0658e4ff7ae07b7cb33f0fb88b9da861 over 1,536,622 bytes.

Release metadata now binds each public SDK to its source repository, immutable commit and tag, registry artifact URL, byte hash, and package-manager integrity metadata. Production gates download those public artifacts and reject source, version, hash, or OpenAPI drift.

Python SDK 2.4.1 and TypeScript SDK 2.3.1 were published from their public repositories at tags v2.4.1 and v2.3.1, the TypeScript package with npm provenance.

The release was tagged api-v1.1.1, keeps MCP at 0.5.8 and applies an additive database migration.

Regression verification covered 1,753 backend unit tests, 239 release/database/graph tests, 40 Python SDK tests, 22 TypeScript SDK tests, 100 deterministic forced-pending single writes without duplicate POSTs, clean public-registry installs, 18 production behavioral contracts, ECS convergence, and a clean post-deploy Sentry window.

World Model remains withdrawn from the public surface. RL training and checkpoint mutation remain admin-only, evaluation and metrics remain plan-gated, and regulated workloads still require the applicable signed agreements.

All public retrieval modes now share one safety envelope. Full, fast, advanced, reason, chat, SDK, and MCP responses expose the same grounded/degraded state, evidence memory IDs, and claim receipt instead of allowing each adapter to invent its own confidence contract.

Open-vocabulary claim grounding now verifies the requested subject, predicate, and value together. Unsupported attributes fail closed even when another fact about the same entity is nearby, while paraphrases of stored facts remain retrievable without fixture-specific aliases.

Cached and uncached retrieval use the same grounding and safety versions. Cache keys include the query normalization, retrieval, grounding, and safety contract versions so a release cannot reuse an answer evaluated under an older policy.

Reason and chat no longer turn related memories into an unsupported answer. When evidence is absent or degraded, the answer is nullable and the response carries an explicit abstention reason.

Corrections suppress superseded wording across retrieval adapters, normalize ASCII and Unicode possessives, and preserve explicit old-to-current provenance, making the corrected fact the canonical current truth without deleting audit history.

The hosted 26-tool MCP v0.5.7 adapter rejects malformed, missing, degraded, or ungrounded upstream evidence, marks returned memory text as untrusted data, and keeps pending writes separate from verified search results.

Python SDK 2.3.0 and TypeScript SDK 2.2.0 add typed scope, readiness, correction, Outcome Memory proof, grounding, idempotency, and fail-closed safety contracts guarded against the exported OpenAPI artifact.

The production acceptance holdout passed 200/200 unseen semantic positives with recall@5 1.00 and MRR@5 1.00, plus 2,000/2,000 subject-match/attribute-mismatch abstentions with no false positives, no false negatives, no retries, and no 5xx or transport errors across eight isolated tenants.

This release makes grounding status canonical on every adapter, including compound known queries and fail-closed degraded responses.

There is no schema or data migration in this release.

Every public memory write now follows one truthful readiness contract. Raw, batch, Mem0-compatible, fast-ingest, memory-tool, correction, and direct-update paths return 200 or 201 only when their result is ready; a durable commit still converging returns 202 with searchable:false, Retry-After, Location, status_url, and a stable indexing event ID.

Memory-tool inserts now forward Idempotency-Key into the PostgreSQL transactional outbox and remove request-time metadata from the logical fingerprint. Identical retries replay one resource and one event; a changed body with the same key returns 409.

Corrections now persist their resolved collection scope, expose GET /v1/corrections/{id} as a real status endpoint, and report canonical search-delivery state from the backing correction memory. Default-collection corrections are discoverable through relevant-search instead of being hidden by a nullable request alias.

Search thresholds are now terminal views over one cached, calibrated score vector. Numeric thresholds no longer create separate retrieval or reranking executions, so raising a threshold can only remove results and cannot introduce a new result.

Open-vocabulary claim grounding now combines deterministic subject binding with strongly calibrated semantic support for label-free paraphrases such as editor transitions, pet breeds, rollback commands, time zones, retry limits, and SLO values, while weak same-entity attribute noise still abstains.

Direct memory edits now preserve immutable history and use the same pollable asynchronous contract when re-indexing is pending, eliminating the previous window in which PATCH returned 200 before search could observe the corrected value.

The generated OpenAPI 3.1 artifact documents every asynchronous 202 response and readiness header, includes correction status GET, and points release traceability to this public changelog instead of a private repository URL.

Compatibility and fast-ingest paths no longer convert bounded post-commit indexing observation into 503 errors. Durable identities remain stable and the outbox reconciler continues delivery without encouraging duplicate client work.

Retrieval now preserves full semantic verification for ambiguous entity and facet queries while safely abstaining before reranking only when dense similarity is below the calibrated unrelated-tail boundary and neither sparse nor exact evidence exists.

Fresh production regression gates cover direct facts, paraphrases, multi-hop questions, distractors, superseded facts, unknown questions, concurrency, idempotent retries, correction readiness, derivative deletion, and OpenAPI/runtime agreement.

Hebbrix MCP 0.5.6 remains the compatible 26-tool public adapter for this API release.

Introduced infer:true mode for automatic fact extraction from conversations. Send raw conversation turns and Hebbrix extracts clean, atomic facts using a 3 tier pipeline: embedding classifier, LLM extraction, and content deduplication.

Added memory worthiness classifier that rejects queries, greetings, and noise before they reach the LLM. Saves token costs and keeps your memory index clean.

Fact extraction now runs on gpt-5-nano for faster, cheaper inference compared to the default model.

New SEARCH_MIN_SCORE setting lets you configure a minimum relevance threshold. Results below this score are automatically filtered out, reducing noise in search responses.

Upgraded the search engine to combine five retrieval layers: semantic vectors, BM25 keyword matching, knowledge graph traversal, importance scoring, and recency boosting.

Added cross encoder reranking option for production workloads that need maximum precision.

Search results now include score explanations showing which layers contributed to each result's ranking.

Improved proper noun and keyword matching through tuned BM25 integration. Queries like 'Sarah from TechCorp' now correctly prioritize exact entity matches.

Entity extraction now runs asynchronously in the background, removing it from the critical path of memory storage. Write latency significantly reduced.

Knowledge graph indexing can be toggled per environment using the ENABLE_KNOWLEDGE_GRAPH_INDEXING setting.

Fixed a timeout issue where Neo4j connection failures would block the entire memory ingestion pipeline.

New /v1/chat/completions endpoint that follows the OpenAI API format. Drop Hebbrix into any existing OpenAI integration by changing the base URL and API key.

Automatic memory retrieval during chat: the system searches relevant memories and injects them into the conversation context before generating a response.

Support for streaming responses with memory augmented context.

Added webhook support for memory lifecycle events: created, updated, deleted, and searched.

Configurable webhook endpoints per collection, with automatic retry and failure logging.

Webhook payloads include a content preview (the first 200 characters), metadata, and event context; deleted memories and document events carry no content.

Added TrustedHostMiddleware to reject requests from unauthorized origins.

Improved API key rotation with zero downtime key swapping.

Added rate limiting per API key with configurable burst and sustained limits.

All memory content is now encrypted at rest using AES-256.

Complete rewrite of the memory storage engine. Memories now flow through three tiers: short term for recent context, medium term for ongoing relationships, and long term for permanent knowledge.

Automatic memory promotion and decay based on access patterns, importance, and recency. Memories that matter get strengthened. Noise fades naturally.

New collection system for organizing memories by application, namespace, or user. Full multi-tenant isolation by default.

Bulk memory operations: import, export, and delete across collections.

New dashboard for managing memories, viewing knowledge graphs, and monitoring usage.

Automatic entity and relationship extraction from stored memories. Store a sentence like 'Alex joined the product team and reports to Jordan' and Hebbrix maps the entities and connections automatically.

Graph traversal during search: ask about Jordan's team and Hebbrix finds Alex through the relationship graph, even without a direct mention in the query.

Visual knowledge graph explorer in the dashboard.