Ghi chú phát hành Organizational AI Memory

Trang này ghi lại các phiên bản Organizational AI Memory đã phát hành. Mỗi mục được tổng hợp từ release note đã qua review, gắn với một phiên bản semantic duy nhất và được công bố cùng bằng chứng build bất biến.

Phần nội dung chi tiết của từng release được giữ nguyên bằng tiếng Anh vì đây là bản ghi chuẩn do Tegami sinh tự động. Các bản gần nhất nằm trên trang này; những phiên bản cũ hơn được chuyển sang kho lưu trữ phát hành và vẫn truy cập được dù repository nguồn là private.

Organizational AI Memory v0.5.0

Clarify the owner-led Asset lifecycle

Documentation

Public guidance now distinguishes mutable working copies, independent sharing state, immutable exact Releases, ownership transfer, Asset withdrawal, and the per-user Skill Enabled preference. It also identifies the optional writable review and deprecation compatibility APIs without implying that the current POC enforces organization-wide review.

Keep Drive permissions current through bounded crawls

Features

Google Drive crawls now bound retained text without abandoning permission checks for the remaining files. Administrators can inspect content and permission health separately, understand partial crawl outcomes, and open the exact crawl limits that need attention from the connection detail page.

Author governed Prompts in the Asset library

Features

Create and edit private Prompt working copies in the Asset library with typed variables, bounded synthetic test cases, and optional Knowledge requirements. Owners can publish an immutable version directly, then run the tests embedded in that exact release without adding a review workflow or changing who can see the Asset.

Organizational AI Memory v0.4.2

Skip release mutation for an idle phase

Operations

A completed release with no newer entry now stops after phase detection. The release workflow no longer invokes versioning or publication for that idle state, while pending recovery and new version work retain their exact existing gates.

Organizational AI Memory v0.4.1

Keep completed releases idle on later commits

Operations

After a product release is fully published, a later green-main change with no release entry now remains an explicit release no-op instead of retrying the old release against newer repository content. Pending releases still resume and continue to block newer version work until recovery.

Organizational AI Memory v0.4.0

Work with private files in Assistant

Features

Upload or reuse up to three private files directly from the Assistant paperclip without publishing them to a Knowledge Space. Private files stay visible only to their uploader, use the same structured document processing as Knowledge, and expire automatically after 30 days. Publishing durable governed Knowledge remains a separate explicit action.

Security

Private-file turns recheck the uploader, organization, lifecycle, exact processing generation, and embedding profile before retrieval or download. Delete and expiry deny access before extracted content and stored objects are cleaned up; old answers retain only a non-clickable unavailable citation marker.

Organizational AI Memory v0.3.0

Keep Assistant evidence useful after reload

Features

Assistant citations now survive transcript reload while still rechecking each user's current access. Source opening starts with the exact governed evidence excerpt, supports safe Markdown, PDF, image, text, and download-only files, and shows truthful retrieval and answer-preparation activity before the first token.

Improve the Assistant conversation experience

Features

The Assistant now restores in-session conversation drafts, offers server-curated starting prompts, retries completed answers with fresh governed retrieval, and lets users save helpful or not-helpful feedback on an answer.

Choose a governed model in Assistant conversations

Features

Choose an administrator-approved model directly from the Assistant composer. The selected model stays with the conversation, remains bound to the exact approved gateway route, and safely falls back to the deployment default when no explicit choice is made.

Use governed Skills in Assistant answers

Features

The Assistant can now discover an authorized Skill, load its exact released instructions, and read bounded supporting text while preparing a grounded answer. Skill content never grants tools or permissions, and OrgMemory does not execute package scripts, binaries, or shell commands.

Restore organization-wide document visibility

Fixes

All-employees demo documents now target the organization-wide Company Knowledge space instead of inheriting a department-only audience. Department and executive documents retain their existing restricted placement.

Make Assistant no-answer guidance permission-safe

Improvements

The Assistant now responds in the user's language when accessible documents do not answer a question, offers one concise next step, labels nearby information instead of presenting it as the requested answer, and cites every source it uses without citing unrelated documents. Assistant answers also show a short reminder that they rely only on documents the user can access.

Keep parallel knowledge graph relations visible

Fixes

The Knowledge graph no longer fails to load when two distinct semantic relations connect the same directed pair of entities. Parallel relations remain separate and visible in the graph and entity inspector.

Keep Knowledge Graph controls from crushing the page title

Improvements

The Knowledge Graph workspace now keeps its page title readable while the explorer controls wrap into the remaining desktop width. The same shared header behavior prevents dense action groups from collapsing page identity elsewhere in the product.

Refine Knowledge document operations and graph inspection

Improvements

Documents now use a persistent desktop reader while keeping the list interactive, show ingestion failures without relying on hover, and guide quarantined evidence into a corrected upload. The graph inspector now presents readable entity context, directional connections, and permission-verified document evidence instead of generic numbered sources.

Complete the governed document reader

Improvements

The employee workspace is now named Knowledge, with Documents and Knowledge graph as its two clear surfaces. Documents open in a responsive right-side reader with safe rendered or raw Markdown, inline PDF and image previews, plain-text reading, explicit download-only fallbacks, and retry when governed content cannot be loaded.

Faster, fairer assistant retrieval under load

Improvements

Assistant knowledge retrieval now admits snapshot queries through one fair process-wide limit instead of per-request batches, so concurrent conversations can no longer exhaust the database connection pool and stall at the turn timeout. The API connection pool is right-sized for the production host, retrieval breadth returns to the upstream LightRAG default, and new payload-free timing stages make the previously unattributed portion of time-to-first-token observable.

Verify retrieval recall before query cutover

Improvements

Operators can now capture and score authorization-preserving retrieval recall against an explicitly restored projection copy without generating answers or touching the live database. The recorded 43-case comparison confirms that the raw-query bypass stays level with the current keyword-seeded path and preserves the evidence needed to diagnose shared misses before any query-plane cutover.

Unify governed document previews

Improvements

Knowledge documents and Assistant citations now open in one centered, responsive viewer. Long PDF, image, Markdown, and text evidence gets the full reading surface, inline citations open it directly, and the source sidebar remains available for comparing cited and discovered evidence.

Rebuild AI gateway consumers for production

Fixes

Production releases now rebuild the API and worker whenever their shared AI gateway integration changes, so approved Assistant and model-routing fixes are included in the immutable image set instead of being treated as deployment no-ops.

Keep Assistant tool calls compatible with OpenAI

Fixes

Fresh production deployments now set Answer reasoning to none so the Assistant's governed Skill tools work with gpt-5.6-sol on OpenAI Chat Completions without requiring an organization route workaround.

Tell people what to do when an Assistant turn fails

Fixes

A failed Assistant turn now ends on a sentence naming what the person who hit it can do next, instead of one generic message for every cause. An expired gateway key, a rate limit, a model that is no longer offered, a gateway that did not answer in time, and a busy assistant are now distinguishable and separately actionable.

Every message remains a fixed sentence chosen from the failure's category, so a misconfigured or unusually talkative AI gateway cannot surface its own text, credentials, or prompt content in the browser. A failure that matches no known category still ends on the previous generic message.

Keep an Assistant conversation in one place

Improvements

An Assistant conversation is now stored once. The transcript that already served history, replay, rename, and delete is also what the model reads back as prior context, replacing a second copy that was kept in a table with no organization, no owner, and no link to the conversation it belonged to.

Prior context is now read in whole question-and-answer turns rather than by counting messages. The question of the turn currently being answered can no longer be sent to the model twice, a turn that failed before answering no longer occupies the window, and the window can no longer begin partway through an exchange. Deleting a conversation removes its context in the same operation instead of relying on a separate call.

Make Assistant prompt controls consistent and accessible

Improvements

The Assistant composer now uses a consistent Prompt Input control set for its model picker, status-aware submit behavior, keyboard composition, tooltips, and future action menus. Text submission, stop controls, and input-method editing remain predictable without enabling file, screenshot, voice, or source attachment capabilities.

Show Assistant Skill activity without a blank wait

Improvements

The Assistant now keeps its progress state visible until answer text appears and shows a compact, current-turn receipt when it successfully activates a governed Skill. Skill titles are bounded plain text, denied or failed Skills remain unnamed, and the receipt clears safely when a turn ends without an answer.

Match Assistant requests against authorized Skills

Fixes

The Assistant now sees the current user's authorized Skill names and descriptions before choosing a workflow, so natural-language requests can activate the matching exact release without inventing catalog search terms. Unavailable Skills remain hidden, and activation still rechecks access.

Keep Assistant activity visible until the answer appears

Fixes

Assistant thinking and Skill activity now stay in one stable transcript position until meaningful answer text appears. Skill receipts no longer expose an empty disclosure when there is no resource detail to show.

Stop losing an Assistant answer that was already on screen

Fixes

An Assistant answer could disappear from a conversation after being delivered. When two turns of the same conversation finished at the same moment, only one of them was saved; the other was rolled back and was gone on the next reload, even though the person asking had watched it arrive. Both are now kept.

Long answers with many sources also render far more cheaply. Each arriving source used to discard and rebuild the entire answer shown so far, which made a long, heavily cited reply progressively slower to display and could leave the page unresponsive while it finished. The answer is now updated in place as its sources arrive.

Separate data clearance from user roles

Improvements

The user "role" field is now a data clearance with two values, Standard and Executive, matching what the system actually enforces: Executive widens confidential and restricted document access, while action permissions stay governed by organization roles. Administrators can now assign a user's department (required for confidential document access), raising someone to Executive asks for confirmation and states its reach, and every user can see their own department and clearance in the account menu. Legacy titles such as Team lead, Manager, Director, and the misleading Admin label are removed; existing Executive users keep Executive and everyone else becomes Standard.

Reuse exact query embeddings across retrieval requests

Improvements

GraphRAG and hybrid knowledge retrieval now reuse exact query embeddings within an explicit projection namespace. Repeated requests avoid duplicate embedding provider work while authorization, evidence selection, and citation verification continue to run normally. Cached vectors remain isolated by embedding profile, provider version, and dimensions, with bounded PostgreSQL retention and expiry.

Keep synthetic redaction fixtures in secret scanning

Fixes

Secret scanning now narrowly recognizes the API-key-shaped value in the assistant redaction regression as synthetic test data while continuing to scan all other files and generic API-key findings.

Keep GraphRAG degree ranking within its retrieval budget

Fixes

GraphRAG degree ranking now resolves authorized relation visibility once and uses indexed source and target endpoint lookups. PostgreSQL also cancels an abnormally slow degree query before the assistant retrieval deadline, avoiding orphaned database work that could degrade later chat turns.

Bound authorized GraphRAG relation loading

Fixes

GraphRAG now resolves authorized relation candidates with set-based entity and relation visibility checks instead of repeating correlated ACL work for each candidate. Relation reads also use the transaction-scoped PostgreSQL timeout so a retrieval cancellation cannot leave database work running in the background.

Restore fast authorized GraphRAG relation scoring

Fixes

GraphRAG now limits relation contribution and relation-weight authorization work to the requested candidate relations before checking independently visible source and target entities. These reads also use the transaction-scoped PostgreSQL budget, preventing an expensive plan from continuing after retrieval is cancelled.

Improvements

Snapshot retrieval now reports bounded, payload-free timings for each graph storage operation under the existing retrieval operation identifier, making future latency regressions attributable without recording prompts, answers, or evidence identifiers.

Filter and page the Documents list

Improvements

Documents can now be narrowed by Knowledge Space and classification alongside the existing status tabs, so the Space shown on every row is finally something you can filter by. Filtering and search run on the server and the list is paged, so a large library no longer arrives in one response. Status tab counts describe the whole filtered library rather than the page you happen to be on, and a document still processing keeps its real status instead of disappearing while it publishes. The knowledge graph search now runs as you type, matching the Documents tab instead of waiting for a separate button.

Keep the knowledge graph stable while its panel is sizing

Fixes

The knowledge graph now waits for a visible, positively sized canvas before starting Sigma, and releases the renderer if the panel becomes size-less during a layout transition. Opening the graph while its tab or flex layout is still settling no longer crashes the page with a zero-height container error.

Put production services on one shared Docker DNS fabric

Fixes

OrgMemory, documentation, and observability services now join the same external Docker DNS network while retaining their existing private and proxy networks. Cross-stack diagnostics and integrations can use stable service names instead of container IP addresses without publishing additional host ports.

Keep the document list working for organization-wide sources

Fixes

The Documents list no longer fails when a source belongs to an organization-wide Knowledge Space. Those sources carry no owning department, and the provenance lookup rejected the missing identifier.

Clarify document provenance and content availability

Fixes

The Documents ledger now identifies each document's Knowledge Space, owning department, and uploader. Published documents outside the current user's content scope now show honest access guidance instead of being described as still waiting for publication.

Enforce the supported Assistant question length

Fixes

The Assistant composer now displays and enforces the 1,000-character question limit before a turn starts. Questions at the boundary remain accepted, while longer input is blocked instead of opening a stream that later fails.

Brand the OrgMemory sign-in experience

Features

• Give OrgMemory sign-in and password recovery a responsive, accessible visual theme aligned with the product's light and dark design tokens.

Operations

• Package the Keycloak login theme in the immutable Keycloak image and reconcile existing realms with rollback-safe restoration of their prior theme.
• Deploy production images by their verified manifest digests. Keycloak rollback now fails closed if the previous realm theme cannot be restored and verified, rather than starting a potentially incompatible previous image.

Make the Asset catalog easier to scan

Improvements

Asset cards now use distinct, consistent marks and prioritize the capability's name, description, human owner, and sharing state. Technical coordinates and shortened owner identifiers no longer compete with discovery, while Skill activation remains available as a compact personal setting.

Share and use Assets from one company library

Features

The Asset catalog now separates reusable releases available to you from Assets you created, and shows each Asset's owner and sharing state. Owners can publish immutable updates directly, share Viewer or Editor access, transfer ownership, and withdraw an Asset without routing new contributions through review. Historical review evidence remains available in the Asset workspace.

Skills now require a personal Use in Assistant opt-in in addition to live access to a released version, so each person controls which shared Skills may guide their Assistant sessions.

Use governed files as Assistant evidence

Features

• Upload up to three supported documents from the Assistant composer, publish them to a chosen Knowledge Space, and wait for governed ingestion before use.
• Keep the exact ordered file selection across a failed retry and cite the same permission-verified evidence used for the answer.

Security

• Recheck conversation ownership, current Source revision, actor access, and active retrieval-engine readiness before each selected-file turn.
• Keep selected files as a hard retrieval ceiling through graph expansion and citation output; direct provider files and transient attachment bypasses remain unavailable.

Deploy releases from a clean checkout

Fixes

Production deployment now executes the exact released commit from an ephemeral clean linked worktree. Staged or local operator changes on the host can no longer replace deployment scripts or image references during rollout.

Ingest the document formats organizations already use

Features

Knowledge document upload now accepts CSV, Excel, legacy Word and PowerPoint, HTML exports, RTF, and OpenDocument files alongside the existing PDF, modern Office, Markdown, and text formats. Upload and processing limits adapt to each format instead of applying one global ceiling.

Document processing now preserves headings, paragraphs, tables, spreadsheet headers, and PDF page provenance through a reusable parser boundary. The versioned structured policy keeps retries deterministic while table fragments retain the header context needed for useful answers.

Preserve rollback evidence after bootstrap cleanup

Operations

Production deployment now tolerates cleanup of the completed PostgreSQL bootstrap container when the previous release is already pinned by exact digest and that same image remains available locally for no-pull rollback. Mutable, missing, or mismatched rollback images remain blocked before rollout.

Organizational AI Memory v0.2.0

Route ACL reads through an owned query boundary

Improvements

Retrieval and Graph now read source ACL facts through SourceAclQuery and immutable ACL-owned references instead of consuming the ACL repository, JPA entity, or a Space-owned projection type. This follows the independently reviewed architecture direction to replace direct implementation edges before closing nested modules.

Remove the ACL-to-Source Ledger dependency

Improvements

ACL heads now consume an ACL-owned source target and preserve stable conflict semantics without importing Source Ledger entities or exceptions.

AI management permission visible in admin, leaner API surface

Improvements

The organization permission catalog now includes the AI management permission, so admin screens can show and explain who may manage AI gateways. The redundant identity endpoint was removed in favor of the session endpoint, and request handling avoids repeated per-request work in SCIM limits and worker scheduling configuration.

Repair published graph snapshots after Apache AGE cutover

Fixes

Production deployment now reconstructs retained relational graph publications in Apache AGE through a bounded, verified one-shot before the API and worker start. Graph exploration and Assistant retrieval no longer remain unavailable when published snapshots predate the AGE topology backend.

Repair Apache AGE startup for least-privilege roles

Fixes

Apache AGE startup now verifies session preload through a bootstrap-owned boolean probe instead of requiring the application to read all PostgreSQL settings. Production-shaped conformance tests use a non-superuser role and keep the broader pg_read_all_settings privilege denied.

Refine Asset catalog layout balance

Improvements

Tighten the desktop Asset catalog with a compact ownership scope, trailing result and layout controls, and clearer selected states while preserving the existing responsive workflow.

Move catalog and chunk values to Knowledge Asset

Improvements

Knowledge Asset now owns its catalog projection, normalized text-chunk value, and PostgreSQL vector encoding, so Retrieval and other consumers depend on the domain that persists and publishes those values.

Add personal Asset ownership navigation

Improvements

Add an owner-scoped My Assets workspace alongside the authorized shared catalog, with responsive search, scope, type, sort, and layout controls.

Remove the Knowledge Asset dependency on Retrieval

Improvements

Knowledge Asset now owns its compact embedding-profile reference and projection namespace identity. Connector and Worker translate richer Retrieval profiles at the orchestration boundary, leaving Asset with no direct Retrieval dependency.

Isolate Asset publication from Source Ledger persistence

Improvements

Knowledge Asset promotion now receives validated source facts through public Source Ledger contracts, while publication advances the source revision through an owner-defined transaction-aware service instead of cross-module repository access.

Unify authorized graph traversal

Fixes

Authorized graph expansion now follows one deterministic core policy across PostgreSQL, Neo4j, and OpenSearch. Exact snapshot validation, permission-scoped paging, canonical ordering, and global limits no longer vary by storage backend, and incomplete native traversal prefixes can no longer become public query results.

Automate release note navigation

Documentation

Generate localized release-note navigation and an internal archive from the reviewed product changelog so public history remains current without manual sidebar edits or repository access.

Author and revise Skill Drafts in the browser

Improvements

Create a private Skill Draft from scratch, a SKILL.md, ZIP, or local folder. OrgMemory previews the validated package before creation, and authorized owners can replace the mutable Draft package without changing any immutable Revision or Release.

Create Skill Drafts from the Assets catalog

Improvements

Open the new Add asset menu from the Assets catalog, choose Skill, and upload a validated ZIP package into an authorized Knowledge Space. A successful import creates a private Draft and opens its existing Governance workspace.

Separate Changelog from documentation categories

Documentation

Keep Changelog as a standalone global documentation surface with focused, automatically generated release navigation while limiting the documentation category switcher to the four reader-oriented categories.

Repair existing-version CLI publication verification

Fixes

Make the retry-safe npm publication path parse correctly when the exact CLI version already exists, and prevent indented nested heredoc terminators from reaching the workflow again.

Correct the missing-version npm publication probe

Fixes

Allow a new exact CLI version to reach npm Trusted Publishing while retaining the immutable-integrity comparison for versions that already exist.

Make CLI publication verification retry-safe

Fixes

Recover a successful immutable CLI publication when npm provenance propagates after the package manifest, while refusing any existing version whose registry integrity differs from the reviewed tarball.

Close the Knowledge ACL module boundary

Improvements

Knowledge ACL now enforces a closed public API with an explicit dependency allowlist limited to organization, permission, shared, and shared::error. This completes the independently reviewed ACL closure after its sibling implementation edges were replaced with owned APIs.

Close the Knowledge Connector module boundary

Improvements

Knowledge Connector now enforces a closed Spring Modulith boundary and an exact outgoing dependency allowlist after its ACL, Source Ledger, storage, Asset, and Retrieval interactions were reduced to intentional public contracts.

Close the Knowledge Graph module boundary

Improvements

Knowledge Graph now enforces a closed public API and an exact outgoing dependency allowlist after Asset, Source Ledger, ACL, Space, and embedding profile persistence access was replaced with owned query and registry contracts. This completes the independently reviewed Graph closure.

Close the Source Ledger module boundary

Improvements

Source Ledger now enforces a closed public API and an explicit allowlist for its ACL, storage, organization, permission, and shared dependencies.

This completes the mechanical closure gate required by the independent architecture review.

Close the Knowledge Space module boundary

Improvements

Knowledge Space now enforces a closed public API with an explicit dependency allowlist limited to authorization, knowledge.sourceledger, organization, permission, shared, and shared::error. This completes the independently reviewed Space closure after sibling repository access was replaced with an owned query API.

Isolate connector read views from Source Ledger persistence

Improvements

Connector inventory and activity views now use a Source Ledger-owned read boundary instead of consuming its repository, entity status, and aggregate projection types directly. The immutable query result preserves active and archived counts, latest activity time, and active external object ids.

Make live connector polling consistent and rotation-aware

Improvements

Slack, Google Drive, and GitHub now run through one connector polling driver for connection isolation, content cadence, failure activity, and credential-derived client lifecycle. Clients retain safe token and rate-limit state across polls, rebuild after credential or client-identity changes, and retire when a connection disappears. A crawl in which at least half of the eligible source units cannot be read is reported as mostly_failed instead of advancing a broadly degraded checkpoint, while all existing crawl and component cursor bytes remain unchanged.

Isolate connector source lifecycle operations

Improvements

Connector reconciliation now resolves source identity, diffs active inventory, and retires tombstoned sources through Source Ledger-owned query and lifecycle APIs. Connector no longer consumes the Source Object repository, entity, or status enum for these flows.

Isolate connector source revision transactions

Improvements

Source Ledger now owns connector revision lookup, evidence staging, completion, and atomic graph-job scheduling behind revision commands and immutable draft facts. Connector no longer consumes Source Ledger repositories/entities or the Graph queue, while independent transaction boundaries remain enforced.

Safer, leaner knowledge publication on OpenSearch

Improvements

Publishing a new knowledge generation on OpenSearch now runs through one coordinated copy-forward protocol with durable ownership: a publisher can no longer take over a copy that another process is still performing, failed copies leave an explicit durable failure state and clean up their partial output, and the previous generation streams across in bounded pages instead of being loaded into memory whole. Stored documents are preserved byte-for-byte.

Keep public docs hydration stable

Fixes

The public documentation now renders its category selector consistently during static generation and browser hydration. English and Vietnamese documentation routes no longer emit React hydration errors when opened from a fresh page.

View and retire governed documents

Features

Let users inspect document metadata, safely view supported document content, and retire eligible manual uploads directly from the Documents workspace while preserving governed evidence and access-control checks.

Recover graph publication safely across worker restarts

Fixes

Graph indexing now binds each cross-store publication to a durable commit permit and claim epoch. Retries resume the exact permitted PostgreSQL and OpenSearch attempt after a worker restart, never discard staging whose visibility is uncertain, fence abandoned copy-forward work, invalidate graph caches before completing the job, and require durable proof before cleaning up a competing attempt.

Explain effective document access

Fixes

Show administrators the OpenFGA relationship decision and the canonical content-policy decision separately, while keeping denied resource metadata behind audit-view permission and technical identifiers out of the primary UI.

Activate evaluated production RAG routes

Improvements

Production now defaults Answer to gpt-5.6-sol, Keyword Planning to gpt-5.6-luna with reasoning disabled, and Graph Extraction to gpt-5.4-mini, preserving the independently evaluated quality and latency split across deployments and shared ZM development.

Select an explicit Apache AGE topology backend

Features

PostgreSQL GraphRAG now selects either Apache AGE or relational topology as an explicit runtime backend. Apache AGE is the production default, fails startup instead of silently falling back, and serves publication-batch-pinned, authorization-filtered topology while PostgreSQL retains canonical evidence.

Import governed Skills from GitHub

Improvements

Discover Skills in a public or administrator-approved private GitHub repository, preview the exact commit and valid packages, then import selected Skills as independent governed Drafts. Private access stays server-side through an approved GitHub App connection, and every Draft retains immutable source provenance.

Accurate review actions in the governance workspace

Improvements

The governance workspace now shows exactly the review actions the server permits: a revision author can request changes or reject, and only approval is withheld, matching the enforcement rule. Action availability comes from the server instead of being derived in the browser, so what you see always matches what you may do. Authorization rechecks behind search, citations, and graph answers now share one hardened implementation with their existing per-surface behavior preserved.

Remove Graph access to Asset persistence

Improvements

Knowledge Graph indexing and curation now consume immutable Asset-owned query contracts instead of Asset repositories, JPA entities, and the chunk projection store. This follows the independently reviewed architecture direction by removing an implementation edge before the Graph module is closed.

Use the LightRAG mini model for graph extraction

Fixes

Default graph extraction independently to gpt-5.4-mini so changing the Assistant model no longer changes indexing behavior, while preserving a dedicated deployment override and immutable processing profiles.

Reliable graph indexing recovery and hardened graph export

Fixes

Cancelling a graph indexing job while it is processing no longer wedges the indexing queue after its lease expires; cancelled jobs settle into a terminal state and the queue keeps flowing. Connector documents whose embedding or publication step failed mid-run now retry cleanly instead of leaving a staged revision that references deleted content. Short knowledge queries with an empty trusted keyword plan no longer fail with an internal error. Graph CSV exports neutralize leading spreadsheet formula characters so exported cells cannot execute as formulas when opened in Excel.

Remove Graph access to embedding profile persistence

Improvements

Knowledge Graph indexing now resolves immutable embedding profiles through the Retrieval-owned registry instead of consuming the profile repository and JPA entity. This follows the independently reviewed architecture direction by removing the final Graph persistence edge before its module-closing cycle.

Remove Graph access to Source revision persistence

Improvements

Knowledge Graph indexing now resolves immutable source-revision state through a Source Ledger-owned query instead of consuming revision repositories, JPA entities, and persistence status directly. This follows the independently reviewed architecture direction by removing another implementation edge before the Graph module is closed.

Faster document ingestion and graph indexing

Improvements

Staged projection writes now travel in bounded batches instead of one statement per row, and the ingestion and graph-indexing workers process a bounded burst of queued jobs per cycle instead of a single job, so backlogs drain far sooner while maintenance jobs and the other queue keep running. Failure behavior, publication atomicity, and stored data are unchanged.

Faster, leaner storage and connector adapters

Improvements

Graph storage adapters now avoid repeated remote round trips on hot paths: vector staging verifies each physical index once per batch, entity degrees are computed in a single pass, and Slack crawls build their member directory once per crawl instead of per thread. Duplicate and dead adapter code paths were removed and model-key handling was hardened, with no change to stored data, cursors, or fingerprints.

Bind independent production AI model routes

Fixes

Production API and worker configuration now bind Keyword Planning independently to gpt-5.6-luna with reasoning none instead of silently inheriting the Answer model from a shared Java default.

Break the ACL-to-Connector dependency

Improvements

ACL now owns its ingestion commands and membership evidence while Connector maps crawl payloads at the boundary, removing the reciprocal module dependency without changing connector or source-access behavior.

Isolate Knowledge Graph lifecycle and processing boundaries

Improvements

Knowledge Graph indexing, processing profiles, lifecycle operations, curation, exploration, and export now share an explicit module boundary, reducing coupling and making future graph changes safer to verify and release.

Isolate retrieval and embedding contracts

Knowledge retrieval contracts

Improvements

Query embedding contracts, embedding profiles, and projection namespaces now share an explicit retrieval module boundary, making provider integration and future retrieval changes easier to verify safely.

Complete the Knowledge retrieval module move

Knowledge retrieval runtime

Improvements

Authorized search, evidence and citation assembly, catalog federation, and retrieval persistence now share one explicit module boundary, making the security-critical retrieval flow easier to inspect and evolve safely.

Finish Knowledge root-package ownership

Knowledge root-package ownership

Improvements

The final source identity, group view, and failure-message types now live with their owning Knowledge modules, leaving the parent package free of domain types and ready for enforced boundary closing.

Record the completed production proof for reliable MCP search

Documentation

The engineering roadmap now records MCP search reliability as shipped after the timeout and result-schema repairs were deployed and verified through a real Claude search_knowledge call.

Restore production AI gateway configuration binding

Fixes

Production API and worker processes now retain the complete configured AI gateway when profile-specific credentials are applied, preventing startup failure after adding gateway capability flags.

Publish the product changelog

Documentation

Publish a localized product changelog in the documentation, generated from the same reviewed Tegami release history used for GitHub Releases.

Security

Harden documentation responses with a content security policy, HSTS, explicit HTML revalidation, and removal of framework disclosure headers.

Route RAG workloads by model role

Improvements

Administrators can now configure independent Answer and Keyword Planning model routes, including an explicitly supported OpenAI reasoning effort. Graph Extraction remains visible as a read-only pinned processing route so changing interactive model settings never rewrites existing document graph history.

Hand governed Skill work to local coding agents

Improvements

Copy bounded instructions into a local coding agent to validate and upload a private Skill Draft, or install one exact released Skill into Claude Code or Codex. Every handoff shows its confirmation boundary, uses the existing OrgMemory CLI and current access, and keeps publication and sharing as separate governed actions.

Remove empty Skill authoring side cards

Improvements

Skill creation, upload, replacement, and GitHub import no longer reserve a large side card for explanatory text. Package details appear only after a real inspection, while the one relevant private-repository note stays beside its input.

Deliver the pinned public Skill CLI handoff

Features

Pin Skill authoring and installation handoffs to @orgmemory/cli@0.1.1, keep the CLI's package, OAuth, and MCP identity versions aligned, and document the same exact npx lifecycle in English and Vietnamese.

Harden the first Skill CLI publication

Fixes

Verify the packed OrgMemory CLI executable before publication and bind npm provenance to the exact repository, so the first public package cannot succeed with a missing orgmemory command or an unrelated source identity.

Complete the governed Skill CLI lifecycle

Improvements

Prepare the OrgMemory CLI for provenance-backed npm distribution and complete the local Skill lifecycle with offline full-tree verification, exact-version updates, verified-only removal, collision-safe target ownership, serialized receipt writes, and crash recovery.

Add target-specific Skill installers

Improvements

Released Skills now offer compact, exact-version installers for Claude Code and Codex. The interface distinguishes verified package integrity and supported installation from runtime behavior that OrgMemory has not certified.

Consistent Skill package metadata validation across surfaces

Fixes

Skill package metadata keys are now validated identically by the CLI and the server, closing a drift where whitespace-only keys could pass one surface and fail another. MCP gateway error handling and CLI package safety helpers were consolidated so the same rules live in one place per app.

Route source ingestion through the ACL facade

Improvements

Source ingestion now uses an ACL-owned transactional facade instead of coordinating ACL repositories and persistence entities directly.

Remove the Source Ledger-to-Asset dependency

Improvements

Source Ledger now validates source provenance before calling its own asset promotion port, while Asset owns promotion persistence and retirement. This removes the reverse module edge without changing ingestion idempotency, security lineage, or publication behavior.

Remove the Source Ledger-to-Connector dependency

Improvements

The current source head projection now belongs to Source Ledger, allowing Connector reconciliation to consume it without creating a reverse module dependency.

Remove the Source Ledger-to-Graph dependency

Improvements

Source publication now schedules graph indexing through a Source-Ledger-owned port, while Graph keeps target validation, profile selection, idempotency, and durable queue persistence behind its adapter.

Remove the Source Ledger-to-Retrieval dependency

Improvements

Source Ledger now depends on its own visibility and embedding-profile contracts while Retrieval implements the governed adapters, removing the reverse module edge without weakening authorization or ingestion behavior.

Remove the Source Ledger-to-Space dependency

Improvements

Source Ledger now uses its own compact Space target port for upload and promotion validation, while Space retains authorization and active-directory policy behind the adapter.

Route Space reads through an owned query boundary

Improvements

Graph and Retrieval now resolve Knowledge Space existence and active status through the Space-owned KnowledgeSpaceQuery API instead of consuming the Space repository directly. This follows the independently reviewed architecture direction to remove implementation edges before closing a nested module.

Enforce typed Knowledge Space audiences

Features

Let administrators choose organization, department, or restricted custom audiences for each Knowledge Space. Managed audiences cannot be silently widened, custom viewers fail closed across PostgreSQL and OpenFGA, and the administration UI explains policy drift without exposing internal identifiers.

Clearer errors, consistent sizes, and a lighter web app

Improvements

Admin AI model screens now surface the real error details returned by the server instead of a generic message. File sizes display with one consistent unit convention across the app, and Skill upload limits read the same on every surface. Source upload no longer applies the confidential-classification department rule to unrelated classifications. The Sources screen loads the knowledge-graph viewer only when its tab is opened, the assistant re-renders far less while streaming, and several unused component kits and dependencies were removed for a smaller bundle.

Organizational AI Memory v0.1.1

Update the Skill CLI command parser

Improvements

Update the OrgMemory Skill CLI to Commander 15 while retaining its existing ESM and Node 24 runtime contract.

Update graph database runtime dependencies

Improvements

Update the Neo4j Java Driver to 6.2 for the graph storage adapter and refresh the JUnit platform used by the backend test suite.

Refresh Node application runtime dependencies

Improvements

Update the MCP CLI, Assistant web client, and public documentation runtime dependencies to their latest compatible minor and patch releases.

Organizational AI Memory v0.1.0

Changelog layout

Documentation

Keep the product changelog title first and render each reviewed release entry with one clear subject heading.

Release evidence secret scanning

Security

Keep full-history secret scanning enabled while recognizing immutable GHCR references whose public image tags contain Git commit SHAs.

Product release management

Operations

Add reviewed product changelogs, semantic versions, immutable artifact manifests, and GitHub Releases without changing the SHA-addressed deployment pipeline.

Shared ZM team development

Operations

Let developers run local OrgMemory applications against the shared non-production ZM services through private SSH tunnels, with automatic worker coordination and guarded post-merge schema/model updates.

Version pull request preflight

Fixes

Allow Tegami to validate a generated Version PR commit while retaining the independent current-main checks around writable release operations.