PowerContext v1.5.22026-11-16
PowerContext v1.5.2
Not only memory, but a full story.
English
PowerContext v1.5.2 makes Agent long-term memory more efficient, transparent, and easier to use. This release improves single-Server performance, introduces native Dashboard observability, expands Desktop and plugin capabilities, improves Gateway request lifecycle handling, and makes cross-tool handoffs more reliable. It also strengthens privacy, write confirmation, and protocol checks across Agent integrations.
More throughput with less CPU work
- Improve write efficiency through batched writes and fewer database round trips.
- Reuse query structures, optimize vector conversion, and reduce repeated computations on the Server.
- Maintainer-provided release measurements report approximately 190% higher write throughput, 171% higher query throughput, and 64%–67% lower CPU cost than v1.5.1.
See batched writes, Scope-bounded SQLite FTS, and bounded related recall.
See processing progress directly in the Dashboard
- Add native views for processing progress, failure diagnostics, Token usage, and runtime state.
- Inspect saved Sources, generated Artifacts, Dream inputs and outputs, and read-only processing status to understand whether material has been processed and where work encountered a problem.
- Preserve unknown save times for historical Sources without a recorded timestamp instead of inventing a historical value.
See Dashboard observability, processing views, and processing failure diagnostics.
Desktop: install, recover, and manage a local Server
- Improve local Server installation, recovery, and management from Desktop.
- Package independent Ops resources for local operations.
- Add Linux systemd --user service integration.
Protocol checks and local setup support do not replace signed-installer or clean-machine platform acceptance. See local Server management, Ops resources, and Linux service integration.
Plugins: build capabilities around your own Artifacts
- Load directory plugins that generate, render, search, and change their own Artifacts. Plugins can declare their own Source types and expose plugin APIs and sandboxed Dashboard pages.
- Support per-plugin databases on OceanBase and seekdb through powercontextplugin<name>.
- Forward a plugin's Artifact API through the Dashboard and limit Scope listing to that plugin's own readable Scopes.
- Improve user-change evidence and consistency between writes and governance checks.
Hot reload, plugin data upgrades, plugin MCP tools, and standalone Source collection are not included. The new directory-plugin framework replaces the previous managed plugin surface; review the upgrade restrictions below before changing an existing deployment.
See directory plugins and plugin storage, access, and consistency.
Gateway: request deadlines, cancellation, and shutdown
- Improve request deadlines, cancellation, and bounded shutdown.
- Prevent Codex retries after deterministic withhold decisions.
Gateway remains alpha and buffers complete responses for inspection. These changes do not provide live token streaming. See request lifecycle handling and Codex retry handling.
More reliable handoffs across tools
- Persist and recover DSH handoff deliveries, expose connection status and update checks, and embed the Dashboard in the sidebar.
- Add the Claude Desktop Code handoff adapter. Native launch stays disabled by default until the receiving environment can be verified.
- Offer optional PowerContext long-term memory for dsh-pet.
See DSH handoffs, handoff delivery service, Claude Desktop Code, and dsh-pet memory.
Safer capture and more reliable Memory processing
- Strengthen privacy handling in DSH automatic recall and capture, request confirmation for OpenCode writes under default permissions, and detect Memory protocol mismatches across Hosts.
- Improve Codex setup, preserve OpenClaw project bindings during capture, and checkpoint in-session auto-compaction.
- Bound Topic Memory related history, preserve the healthy processing prefix before a failing or oversized look-ahead, and prevent concurrent head conflicts from consuming window attempts.
- Keep Memory merges within the tag limit and allow portable Source exports to retain recorded absolute working directories.
See DSH privacy, OpenCode write confirmation, protocol checks, and Topic Memory retrieval.
Install and upgrade
``bash
uv tool install --force "powercontext[cli,server]==1.5.2"
`
Upgrade Server, SDK, CLI, and integrations together. Use powercontext-v1.5.2 for Agent integration refs and reopen Agent sessions after upgrading.
Compatibility changes:
- The target database revision is v150005. It removes the previous Artifact plugin-framework storage and adds nullable pcartifacts.metadata. Immutable schema revision names retain their original numbering independently of the package version.
- Databases retaining old plugin registrations or data have no supported in-place upgrade path. Migration refuses these databases. Unloading plugins, deleting plugin directories, or exporting archives does not clear retained framework records, even when no Artifacts were generated. Keep the database on the previous version until a supported migration is available. A separate fresh database can use v1.5.2, but this does not migrate old plugin data. Do not manually delete framework records to bypass the check.
- The removed plugin surface includes /v1/admin/artifact-plugins/, powercontext.artifactplugins, and POWERCONTEXTSERVERARTIFACTPLUGINSDIR. Use POWERCONTEXTSERVERPLUGINS_DIR for directory plugins.
- Portable Artifact revision records use schema version 3; restoring version 2 records remains supported.
- Update imports from powercontext.server.authz to powercontext.builtin.authz, from powercontext.cli.envfile to powercontext.envfile, and from powercontext.cli.inferencenotice to powercontext.server.inferencenotice.
- powercontext.paths.defaultschedulerpath has been removed. The scheduler_path argument is deprecated and ignored; passing a non-None value emits DeprecationWarning`. Remove it from runtime construction calls.
For supported migration paths, stop writers, back up the complete database and configuration, inspect the migration plan, and restart only after full Server verification succeeds. Database upgrade support is limited to source-version, target-artifact, backend, and configuration combinations that pass full Server acceptance.
See the database migration guide, [plugin guide](https://github.com/oceanbase/powercontext/b
PowerContext v1.5.12026-11-09
PowerContext v1.5.1
Not only memory, but a full story.
English
PowerContext v1.5.1 makes Memory evidence easier to inspect and deployments easier to operate. This release adds versioned Source authority declarations, precise citation positions, an optional monitoring stack, and expanded Desktop observations. It also improves Memory processing and code-context reuse, strengthens the alpha Gateway, adds subscription-backed generation, and simplifies installation and Agent integration.
Trace Memory back to its evidence
- Retain Source authority and verification declarations alongside exact Atomic Memory revisions. Trusted adapters and deployment configuration can supply these declarations; model output and Source text cannot assign themselves trust. Evidence remains inspectable after restarts, registration changes, and projection rebuilds.
- Preserve neutral untrusted / unknown values for historical records without accepted declarations. The declarations do not change retrieval scores or grant access, and upgrading does not retroactively attest old evidence.
- Add optional citation locators for Source journal positions or exact text line ranges, with an optional heading path. Preserve them through Artifact lineage, Handoffs, HTTP/MCP reads, and portable export/restore. Existing references without locators remain valid; historical locations are not guessed or backfilled.
See Sources and capture.
Operate with connected metrics, logs, and traces
- Add an opt-in Docker Compose monitoring stack with Prometheus, Grafana, Loki, Tempo, OpenTelemetry Collector, and Phoenix, including dashboards, alerts, retention settings, and runbooks.
- Manage it with powercontext observability up, status, check, and down, or use connect to prepare integration with an existing platform. Start the Server with --observability to load the selected telemetry configuration.
- Inspect service health, Scope activity, processing progress, and provider calls, with links to related logs and traces when available. Preserve Worker failure evidence through policy blocks, and expose clearer Topic Memory failure stages, reasons, and budget diagnostics.
The managed stack requires Docker Compose and the tracing-otlp extra. Monitoring is optional; operational telemetry excludes prompts, evidence bodies, model responses, and credentials. See the monitoring guide and processing diagnostics.
Less repeated processing, more faithful Memory
- With built-in prompts, compare earlier candidates from the same Source window only when they meet the related-memory threshold. Candidates with no related evidence can skip reconciliation; vector profiles incur additional embedding work for this selection.
- Reject automatic revisions that repeat the existing kind and text, and recover from exhausted correction attempts for that specific rejection without indefinitely blocking later Sources. Preserve supported explicit effective dates and event times during extraction and reconciliation, including dates adjacent to Chinese text.
- Reuse prepared code evidence with per-file validation. Editing an unrelated file does not invalidate reuse; changing or removing a cited file requires fresh context. Gateway users can opt in with PCGATEWAYINJECTION_CODE=on, with at most one refresh per request when a replayed context becomes stale.
- Use native minimum-term filtering on supported OceanBase backends to reduce query construction and filtering overhead. Retain complete-word coverage on unsupported engines and queries that require the compatibility path. Read Task Outcome receipts by exact reference instead of scanning unrelated results.
See reconciliation changes, temporal evidence, code reuse, and OceanBase retrieval. Performance gains depend on the workload and retrieval path.
Desktop: guided first use and clearer observations
- Guide users from Overview through connection setup, explicit activation, exact Scope selection, and the Memory workspace. Keep recovery actions and operation availability visible throughout the flow.
- Add an Observation area for Runtime state, current-Scope inventory and usage, and this session's requests and actions. Show bounded diagnostics with explicit refresh, reset, and safe preview/copy controls.
- Separate lightweight connection identity checks from explicit readiness probes. Users with Scope permissions can keep using authorized Memory operations when they lack Server observation permission; routine connection checks do not probe models.
- Add versioned candidate assembly and a separate publication workflow that binds installers, Server resources, compatibility records, and checksums to platform qualification evidence.
Desktop remains a preview. Signed distribution and installed-platform qualification are separate from these workflow changes. The first-use guide connects to an independently prepared Server; in-app installation, repair, and service management remain unavailable. See Desktop.
PowerContext Gateway: protocol compatibility and operating limits
- Preserve supported native Chat fields and developer messages, accept Responses string input, and replay tool-argument, reasoning-summary, and citation events after release checks.
- Require PCGATEWAYAPI_KEY for non-loopback listeners and authenticate model requests before reading their bodies. Server authentication also rejects unauthenticated requests before body consumption.
- Check the final rendered request against Core's advertised capacity and the current 1 MiB ceiling. Cap buffered provider responses at 1 MiB; oversized responses are withheld without automatic resend.
- Export bounded, content-free JSONL call evidence to stdout, with operational tracing on stderr. Distinguish provider attempts, reported usage, release authorization, and HTTP-framework observations; retain unknown usage when it cannot be established.
Gateway remains alpha and buffers complete responses before replaying them to clients. It does not provide live token streaming. The shared lifecycle records deadline and cancellation intent; automatic end-to-end deadline enforcement and disconnect cancellation remain separate work. Gateway JSONL also needs its own collector configuration. See the Gateway guide and call evidence.
Reuse existing Codex and GitHub Copilot subscriptions
- Configure openai-codex:<model> with an existing file-backed Codex CLI login, or github-copilot:<model> with an existing OAuth token and a Chat Completions model.
- Use these providers for Memory, Experience, and Profile generation. Codex credential renewal stays with the CLI; Server and Workers read its latest credentials for each request without consuming refresh tokens or modifying the login file.
Topic Memory, Dream, and built-in inference protection are not supported with these subscription providers. Credentials must be available to the Server; a client-side login is not transferred automatically. See provider setup and limits.
Broader Agent integration and capture cont
PowerContext v1.5.02026-11-03
PowerContext v1.5.0
Not only memory, but a full story.
English
PowerContext v1.5.0 introduces Source and Artifact plugins, making data ingestion and custom Artifact definition, generation, search, and management extensible. This lays the groundwork for future multimodal data sources and Artifacts tailored to different business needs.
Token efficiency engineering and inference safety are also key improvements: reduce repeated evidence in model inputs, reuse prepared context, and check final requests and complete responses before sending or delivery. This release also adds clearer evidence for context selection, the alpha Rust PowerContext Gateway, improved Desktop and managed local Server workflows, and historical Server storage migration tooling.
Source and Artifact plugins: extend data ingestion and Artifact management
- Add trusted local Source plugins for synchronous writes and continuous collection. Agent hosts can discover plugin capabilities and manage collection targets through MCP; dedicated collection runtimes retain checkpoints and lifecycle state.
- Define and generate custom Artifact types for business workflows, with versioned schemas, exact Source lineage, and declared text or vector search. Included incident-report and ticket-summary plugins demonstrate how project data becomes reusable, searchable results.
- Administrators can load, enable, disable, reload, and unload Artifact plugin bindings. New Artifact bindings start disabled; saved Artifacts remain readable when generation is disabled or plugin code is removed.
Plugin loading requires deployment administration. Local collection process management currently requires Linux. See the Source plugin guide and Artifact plugin examples.
Token efficiency engineering: smaller inputs and less repeated work
- Reduce redundant model input with reversible shared-JSON encoding. Repeated evidence text, metadata, and structure can share one representation while preserving Source identities, order, repeated occurrences, and data types. Encoding covers Atomic Memory extraction/reconciliation and Experience generation/incubation without additional model calls.
- Select encoded input only when it restores the original stage JSON exactly and the complete representation, including decoding instructions, is smaller in both bytes and estimated input tokens. Otherwise, retain the original input. Atomic Memory input budgeting uses the same representation that generation sends.
- Reuse prepared Atomic Memory within a single Scope using short-lived proofs. Valid reuse avoids repeating retrieval, query embedding, and context assembly while rechecking current permissions and evidence. This saves preparation work; the retained context still forms part of the model input.
- Measure the tradeoffs with per-call encoding estimates and an off / shadow / active evaluation workflow covering model quality, actual provider usage, latency, failures, and unknown usage. Local estimates remain separate from provider token accounting.
Encoding is off by default. shadow measures candidates while sending original inputs; experimental active selects eligible smaller inputs. Savings depend on the workload and model, and exact restoration alone does not establish equivalent model understanding or lower total task cost. See input encoding, context reuse, and the encoding evaluation workflow.
Inference safety: check before sending and before delivery
Protection is opt-in and disabled by default. With a supported target configured, enforcing mode provides the following checks; shadow and advisory observe requests without blocking them.
- Inspect final SDK-rendered model requests before transmission and complete responses before delivery at the owned provider transport. Check supported text and tool arguments against configured sensitive values and optional bounded email, private-key, and API-token rules. Admitted content remains unchanged.
- Verify the declared endpoint, model, protocol, context window, and output limit. Unsupported targets, opaque content, or incomplete inspection are refused or withheld in enforcing mode.
- Let external executors check final payloads through the public context admission contract. Exact Atomic Memory reuse revalidates current identity, Scope, permissions, revisions, and dependencies; forgotten memories, revoked access, or expired proofs stop reuse. Executors must honor the admission decision and preserve the admitted payload.
- Keep inference protection in force during input encoding. Check original content, the selected representation, and the complete final request; a policy refusal cannot be bypassed by falling back to unencoded input. Reversible encoding is not privacy redaction.
- Preserve retryable work after mandatory refusals. Refused Memory write-gate or Topic Memory generation work does not publish its result or consume its Source window; saved Sources remain pending for retry.
See inference protection and context admission and reuse.
Understand context selection and inspect memories
- Request include_receipt during context preparation to inspect selected exact revisions, retrieval and fallback modes, omissions, configuration, and stage timings. The diagnostic Receipt contains no query or evidence bodies. If a complete Receipt cannot be returned, the prepared context remains usable.
- Browse a bounded, paginated Atomic Memory manifest before reading selected exact revisions. New memories retain short descriptions; historical descriptions may remain absent, and browsing the manifest does not generate summaries.
- Preserve the subject, scope, and applicability of extracted facts when Sources describe different entities. Automatic host recall keeps its default response shape without requesting a Receipt.
See context preparation and the Receipt API migration notes.
PowerContext Gateway — alpha
- Add an optional Rust Gateway with OpenAI-compatible /gateway/v1/chat/completions and /gateway/v1/responses endpoints. The Gateway delegates request checks before sending and complete-response checks before release to PowerContext Core.
- Opt into Atomic Memory injection for Chat Completions. Capture turns that pass release checks as Sources for background extraction and later recall; capture is available for both endpoint formats. Injection and capture are disabled by default.
- Run the Gateway independently over TCP or alongside Core over a private Unix domain socket. It is built and started separately from the Python Server.
Gateway is an alpha product. It buffers complete responses before release and does not provide real-time token streaming to clients. Its APIs and behavior may change. See the Gateway guide.
Safer Desktop and local Server operations
- Guide connection activation into exact Scope selection and enable actions according to discovered Server compatibility. Recheck Server and Principal identit
PowerContext v1.4.12026-10-26
PowerContext v1.4.1
Not only memory, but a full story.
English
PowerContext v1.4.1 reduces repeated work in context retrieval and Memory writes, lowers extraction input overhead, and adds portable archives. This release also expands Kimi Code workflows, improves Desktop save recovery, and introduces a complete Server migration framework with explicit upgrade acceptance requirements.
Faster recall and Memory writes
- Consolidate context preparation queries, batch authorization and evidence reads, and reuse unchanged search index text. Dashboard lists, Memory search, and statistics perform less repeated database work.
- Batch Source capture and Memory publication while preserving Source order, exact revisions, permissions, and atomic publication. Failed processing windows leave no partial publication or cursor advance.
- Remove duplicate Source text from extraction input while retaining evidence and provenance. Skip reconciliation when a candidate has no comparable memories, and return invalid reconciliation decisions to the model for correction within the configured retry budget.
- Skip already-consumed Scope windows during scheduled processing. In the built-in relational processing path, new Sources can be captured while Memory extraction waits for the model.
Selected SQLite measurements illustrate the reduction in database work:
| Operation | Before | After |
| --- | ---: | ---: |
| Experience list, approximately 3,000 records per family with authorization | 9,033 SELECTs | 4 SELECTs |
| Capture 3,000 Sources | 75,034 SQL statements | 24 SQL statements |
These are separate, fixed-workload comparisons from read-path optimization and write batching, not a release-wide latency guarantee. Four fixed extraction-input cases also showed 35.7%–59.6% fewer estimated input tokens; this measures serialized extraction input, not provider billing or total Agent cost. Extraction input measurements.
Portable archives with preserved history
- Export selected complete Scopes as .pcb archives, inspect their contents, and check target compatibility and conflicts with a restore dry run.
- Preserve exact identities, immutable revisions, evidence, lineage, Handoffs, and supported review state. Repeated restore recognizes identical records and rejects conflicting immutable content.
- Use powercontext archive export, inspect, and restore for offline administration. Stop writers before restore; credentials, access bindings, and derived indexes are excluded.
Portable archives complement complete database backups and do not replace schema migration. Archive guide.
Inspect blocked processing and recover explicitly
- Inspect durable Topic Memory blocks through HTTP, the Python SDK, or powercontext topic-memory status, block, and retry. Authorized retries retain exact receipts across restarts; acceptance schedules work without guaranteeing generation success.
- Reject semantically invalid Topic Memory output. Configure the recall Embedding time budget and inspect why context preparation fell back from semantic retrieval.
- Optionally aggregate streamed model output to reduce gateway idle-timeout failures. Enable POWERCONTEXTSERVERINFERENCEGENERATIONSTREAMING=auto for supported providers; streaming remains disabled by default and requires a streaming-capable endpoint.
See Topic Memory recovery and streamed generation.
More complete Agent workflows
- Kimi Code: bind working directories to Scopes and opt into recommended or full tool profiles for Handoff, Experience, managed Skill, and Candidate workflows. Initial setup remains minimal; Server capabilities, permissions, and explicit review decisions still apply.
- Add the powercontext scope command group for Scope management and workspace/session bindings. Improve Agent setup recovery, Codex executable discovery, Windows bootstrap receipts, and bounded WorkBuddy Scope discovery.
- Guard automatic prompt capture against recognized credentials and bound long-prompt recall queries. Add an optional Jev reranker for explicit injection by Python applications; it requires separate credentials and is not enabled through Server environment settings.
Desktop and Dashboard improvements
- Add an unsigned macOS app and DMG preview, fix Linux packaged navigation, and show the active connection, Server address, and exact Scope more clearly.
- Open saved Memory records at their exact revisions, preserve drafts across failed or ambiguous saves, and improve permission recovery, confirmations, and keyboard focus.
- Add Experience text search to the web Dashboard and an Observation navigation entry to Desktop.
Desktop remains a preview that connects to a compatible standalone Server. The Observation entry does not yet provide a complete observation-data view; signed distribution and complete installed-platform qualification remain separate work.
Operations and reliability
- Add managed rotating log files with retention and compression, aggregate worker logs through their owning Server, and expose dependency and processing observations. Logs remain best-effort.
- Add the independent powercontext-ops entry for inspecting, configuring, and managing an already-prepared personal Server runtime. Runtime download, repair, and the Desktop installation bridge remain separate work.
- Bound HTTP request bodies and active responses, authenticate MCP calls per HTTP request, and preserve path authorization. Commit supported local writes and ownership records atomically, and reject blank text Sources before they can block later processing.
Install and upgrade
After the package is published:
``bash
uv tool install --force "powercontext[cli,server]==1.4.1"
`
Retain the extras your deployment uses, such as code or seekdb. Upgrade Server, clients, CLI, MCP, and Agent integrations together, then reopen Agent sessions.
The Server migration framework now provides powercontext server db-migrate with status, plan, apply, and verify subcommands. It recognizes registered source layouts, preserves source history and recovery evidence, and validates the full Server before reporting readiness. Migration steps display as v141001 through v141003, while existing revision IDs and checksums remain intact.
Existing databases require an accepted upgrade path for the exact target artifact. The 1.3.0 acceptance matrix declares 12 backend/configuration combinations: publication requires all 10 automatic cases to pass and the support declaration to be reviewed. The two embedded seekDB configurations with Code enabled remain manual-only, unverified, and excluded from supported upgrades. The registered 1.4.0 bridge is outside this acceptance matrix.
Stop all writers, back up the complete database and configuration, and review the migration plan before applying it. Drain queued or running 1.3.0 Dream tasks and preserve the historical Embedding configuration. Follow the documented recovery procedure after interruption, and restart services only after full Server verification succeeds. Database migration guide.
Compatibility changes:
- Custom Atomic Memory and Topic Memory indexes must support complete-set projection publication.
- The old /dashboard/guide route, Notes entry parameter, and Methods kind=all` compatibility paths have been removed.
[Full changelog](https://github.com/oceanbase/powercontext/compare/po
PowerContext v1.4.02026-10-19
PowerContext v1.4.0
Not only memory, but a full story.
English
PowerContext v1.4.0 helps project knowledge evolve while keeping work and changes traceable. This release introduces explicit Dream reviews, Working Notes for current progress, Dashboard revision history and comparison, and an experimental Kimi Code integration.
Dream: evolve knowledge through review
- Use task outcomes, explicit corrections, or usage feedback as a reason to start Dream. Select the relevant knowledge and evidence, then inspect the proposed changes.
- Propose revisions to Atomic Memory, Profiles, Topic Memory, Handoffs, Prompt configuration, and tags through the unified Candidate review flow.
- Approve the exact Candidate version against the current target before publication. Generating a proposal does not approve it, execute work, or install a Skill.
Available operations depend on Server configuration. Skill revision remains disabled until its trusted evaluation requirements are met. Approving a Handoff publishes a revision without automatically continuing the task.
Working Notes: retain current progress
- Capture findings, abandoned approaches, validation results, unresolved questions, and next steps as short-lived Working Notes.
- Cite still-valid notes in a Handoff through exact Source references, so later work can inspect the evidence behind the recorded progress.
- Set optional expiry to prevent new consumption while retaining the original record for audit.
Working Notes stay outside automatic Memory, Topic Memory, and other Artifact generation. Capturing a note does not automatically promote it into durable knowledge.
Knowledge changes with traceable history
- Browse revision history, open exact versions, and compare content directly in the Dashboard.
- Restore supported historical content as a new revision, recording the source version, reason, and actor. Previous revisions remain readable.
- Inspect each Atomic Memory’s history and compare its revisions in a read-only view. Atomic Memory restoration follows its separate preview and confirmation workflow.
New revisions record creation timestamps; existing records retain unknown metadata where historical values are unavailable.
Connect Kimi Code to project memory
- Install and diagnose the experimental integration through powercontext setup kimi-code and powercontext doctor kimi-code, using an explicitly configured Scope.
- Save and search memories, read exact revisions through native MCP tools, and receive relevant context in a new session.
- Optionally capture text prompts as Sources and inspect content-free Hook runtime observations.
Explicit Memory operations do not require a Generation model. Automatic Memory generation requires the Server’s Generation configuration and scheduler; successful Source capture alone does not establish that a Memory was created. Open a new Kimi Code session after setup or configuration changes.
Context, retrieval, and Agent integrations
- Codex and Claude Code: Add opt-in, bounded startup context from explicitly selected Atomic Memories and an optional exact Handoff. Delivery receipts support deduplication; the feature is disabled by default.
- Retrieval: Unify Chinese segmentation and English word analysis across supported backends, configure full-text/vector fusion weights for Atomic Memory and Topic Memory, preserve Source language during generation, and improve long-query handling.
- DSH: Expand native MCP workflows with tool-catalog checks, Scope enforcement, approval handling, and caller-content secret screening. Refresh generated Agent guidance while preserving local edits.
Reliability and evaluation
- Standardize operational logs, metrics, and trace identity, and expose extraction diagnostics through CLI capabilities.
- Preserve exact captured Source identities on reads, pin SQLite content before custom authorization, and improve connection cleanup during cancellation.
- Add stable Scope-local handles for exact Artifact revisions. Omit the placeholder x-api-key when Anthropic-compatible requests use an explicitly configured Authorization header.
- Add cross-host work-continuity acceptance, offline Memory write-gate evaluation, and Dream generation baselines; extend Kimi Code and ZCode acceptance workflows. These checks do not establish a universal quality or performance gain.
Install and upgrade
After the package is published:
``bash
uv tool install --force "powercontext[cli,server]==1.4.0"
`
Include any extras your deployment already uses, such as code or seekdb. Upgrade the Server, clients, CLI, MCP, and Agent integrations together, then reopen Agent sessions.
Back up the complete database and configuration before maintenance. Stop all writers before database changes. Existing collection-based Memory databases require offline Atomic Memory conversion and verification.
Compatibility changes:
- Candidate APIs move from /v1/artifact-candidates/ to /v1/candidates/ without compatibility aliases. Dream clients must support contract 2.
- Retained legacy Memory routes return Atomic Memory records and references. Collection-based operations and legacy citation formats are no longer supported where documented.
- The legacy collection MemoryWriteGate is unsupported by the Atomic Runtime; legacy collection capacity and compaction settings do not constrain Atomic Memory.
Existing Dream database upgrades remain limited. The current unified migration bundle reports serverready=false and does not provide a complete production upgrade path. Passing Atomic Memory verification alone does not establish full Server readiness. Do not bypass migrationrequired or migrationframeworknot_ready.
[Full changelog](https://github.com/oceanbase/powercontext/compare/powercontext-v1.3.0...powercontext-v1.4.0)
---
中文
PowerContext v1.4.0 让项目知识持续演进,也让工作进展与知识变化有据可查。本次带来显式发起的梦境复盘、承接当前进展的便签、Dashboard 历史版本浏览与对比,以及试验性 Kimi Code 集成。
梦境,触发知识演进
- 把任务结果、明确纠正或使用反馈作为复盘契机,显式发起梦境。选择相关知识与证据,检查提出的修订建议。
- 通过统一 Candidate 审核流程,为 Atomic Memory、Profile、Topic Memory、Handoff、Prompt 配置及标签提出修订。
- 批准时检查精确候选版本与当前目标,再发布变更。生成提案不会自动批准、执行任务或安装 Skill。
可用操作取决于 Server 配置。Skill 修订在满足可信评测要求前保持关闭。批准 Handoff 会发布新版本,不会自动续接任务。
便签,留住当前进展
- 用短期便签记录当前发现、放弃的方案、验证结果、未解决问题和待办。
- 在 Handoff 中通过精确 Source 引用使用仍有效的便签,让后续工作能够查看进展背后的依据。
- 可设置过期时间;过期后阻止新的消费,原始记录仍保留供审计读取。
便签不参与自动 Memory、Topic Memory 或其他 Artifact 生成。采集便签不会自动将其晋升为长期知识。
知识变更,有据可回溯
- 在 Dashboard 浏览修订历史、打开精确版本,直接对比内容变化。
- 将支持的历史内容恢复为新修订,记录来源版本、原因和操作者,已有历史仍可读取。
- 查看每条 Atomic Memory 的历史并对比版本,界面保持只读;Atomic Memory 恢复使用独立的预览和确认流程。
新修订记录创建时间;旧记录缺少历史元数据时保留未知状态,不补造日期。
让 Kimi Code 接上项目记忆
- 通过 powercontext setup kimi-code 和 powercontext doctor kimi-code 安装、诊断试验性集成,使用显式配置的 Scope。
- 通过原生 MCP 工具保存和搜索记忆、读取精确版本,并在新会话中注入相关上下文。
- 按需将文本提示采集为 Source,并查看不含正文内容的 Hook 运行观测。
显式记忆操作无需 Generation 模型。自动生成记忆需要 Server 的 Generation 配置和调度支持;Source 采集成功本身不代表记忆已经生成。安装或调整配置后,请重新打开 Kimi Code 会话。
上下文、检索与 Agent 集成
- Codex 与 Claude Code: 新增可选的有界会话开场上下文,使用显式选定的 Atomic Memory 和可选的精确 Handoff。通过投递回执支持去重,默认关闭。
- 检索: 在支持的后端统一中文分词与英文词语分析,为 Atomic Memory 和 Topic Memory 配置全文与向量融合权重,生成时保留 Source 语言,并改进长查询处理。
- DSH: 扩展原生 MCP 工作流,检查工具目录、执行 Scope 约束、保留审批流程,并筛查调用方内容中的敏感值。更新自动生成的 Agent 指引,同时保留本地修改。
稳定性与评测
- 统一运行日志、指标与追踪身份,在 CLI capabilities 中提供提取诊断。
- 精确读取已采集的 Source,在自定义授权前固定 SQLite 内容快照,并改进取消时的连接清理。
- 新增 Scope 内指向精确 Artifact 版本的稳定句柄。为 Anthropic 兼容请求显式配置 Authorization 请求头时,不再附加占位 x-api-key。
- 新增跨 Host 工作连续性验收、Memory 写入门控离线评测及 Dream 生成基线,扩展 Kimi Code 和 ZCode 验收工作流。这些检查不代表所有任务都有统一的质量或性能提升。
安装与升级
包发布后执行:
`bash
uv tool install --force "powercontext[cli,server]==1.4.0"
`
保留当前部署使用的 code、seekdb` 等扩展,一起升级 Server、客户端、CLI、MCP 与 Agent 集成
PowerContext v1.3.02026-10-12
PowerContext v1.3.0
Not only memory, but a full story.
English
PowerContext v1.3.0 makes project knowledge easier to manage and reuse. It introduces independent Atomic Memory, versioned decision policies, Artifact search and shared merge/restoration workflows, a Windows Desktop preview, and an experimental Dify integration.
Independent Atomic Memory
- Store each fact, preference, or decision as an independent atomic-memory Artifact with immutable content revisions, exact evidence references, and a separate lifecycle.
- Create and revise explicit memories without a generation model. Search active memories using text, vector, or hybrid retrieval; vector retrieval requires a compatible Embedding configuration.
- Forget a memory while retaining its content and history. Merge related memories and restore content with explicit revision and state checks. Retired memories cannot be restored directly.
- Convert existing Memory collections through an offline migration that preserves entry version chains and converts exact entry references. Legacy collections move to an offline archive.
Decision models with explicit policies and observations
PR #1921 adds shared contracts for versioned decision policies, typed assessments, and observations, and connects them to Atomic Memory processing.
- Select disabled, shadow, advisory, or enforcing behavior for the Memory evidence gate. Shadow records an assessment while allowing the write; advisory flags insufficient evidence; enforcing can hold a write under the configured gate rules.
- Distinguish substantive judgments from fallback or unassessed results. Observations retain policy/version, evidence references, coverage, final action, latency, and usage without retaining raw candidate or evidence content.
- Make the privacy boundary explicit. The default noexternalcall boundary makes no decision-model call; local_only supports a controlled in-process or loopback backend. Hosted redaction is not enabled by this release.
Model failures preserve the write path. Decision assistance and the gate remain opt-in. When enabled, the gate defaults to shadow with noexternalcall: writes proceed without a decision-model call. Calibrate and evaluate the policy before enabling enforcement. Durable observations are isolated by Scope; this release includes no public audit dashboard. The current integration does not deliver a general replay/rescore runner, automatic policy promotion, or gates for every Artifact family. Existing authorization and candidate approval requirements remain in force.
Find and organize reusable Artifacts
- Search current Experience, managed Skill, Topic Memory, and Atomic Memory through a family-scoped route within a Scope. Results preserve complete content and exact revision references.
- Experience and managed Skill support text search; Topic and Atomic Memory also support configured vector/hybrid retrieval. Controls and score availability differ by family.
- Use shared merge and restoration services for Atomic Memory and Experience, including Atomic Memory restoration previews and stale-state checks. These are explicit operations; writes with an unknown outcome require state inspection before retrying.
- Choose automatic, full-text, semantic, or hybrid Memory search in the web Dashboard and inspect the executed mode, matching channels, and retrieval scores. Local Dashboard access can omit authentication; the Dashboard still requires explicit enablement.
Windows Desktop preview
Connect the unsigned Windows Desktop preview to an existing Server, select a Scope, save notes, search with full-text retrieval, read exact historical revisions, and copy content or references. The preview supports Chinese/English, themes, and connection diagnostics. Complete native interaction and formal release qualification remain unfinished; it does not replace Server deployment.
More ways to connect and troubleshoot
- Dify: An experimental source plugin provides 19 HTTP-backed tools for context preparation, Memory, Source capture, Handoff, Experience/Skill generation, and read-only candidate inspection. Deploy Server separately and configure a fixed Scope. Marketplace publication and complete installed-workflow acceptance remain separate from source availability.
- Codex: Import historical user prompts as ordinary Content Sources, with dry-run and persistent checkpoints. This imports user prompts, not complete conversations or verified conclusions.
- Server: Discover HTTP/MCP contracts and read a persistent deployment identity through /v1/server-info. Read-only extraction diagnostics help distinguish configuration, model, Worker, and processing failures without exposing raw prompts or credentials.
- Agent integrations: Improve DeepSeek Harness Desktop setup and native MCP diagnostics, ZCode workflows and lifecycle hooks, stale endpoint repair, and loopback transport handling.
Reliability and evaluation
- Fix recall-sufficiency decisions to use vector relevance rather than fused ranking. Optional RecallEffort observations are persisted in daily aggregates; the recall gate remains disabled by default.
- Improve extraction timeout recovery and exhausted-window handling, Topic Memory publication authorization, Scope binding errors, and SQLite vector-result limits.
- Extend paired PowerContext OFF/ON continuation evaluation to Claude Code, OpenCode, and Pi; add task-completion and MemoryCode workloads, interval/latency/usage reporting, and LoCoMo-Plus Jev comparisons. Evaluation coverage does not establish a universal quality or performance gain.
- Add compact Handoff continuity hints and Experience/Skill applicability-selection examples.
- Provide a limited unified database migration bundle covering four registered Artifact tables. It reports server_ready=false and is not a complete production Server upgrade path.
Install and upgrade
After the package is published:
``bash
uv tool install --force "powercontext[cli,server]==1.3.0"
`
Preserve the extras used by your deployment, such as code or seekdb. Upgrade Server, clients, and Agent integrations together.
Existing Memory databases require maintenance before the new service starts. Back up the complete database and configuration. Complete any required processing migration first, then use the new CLI and deployment configuration:
`bash
powercontext server atomic-memory-migrate --action plan --env-file .env
Stop all APIs, hosts, Workers, automatic restarts, and input writes before apply.
powercontext server atomic-memory-migrate --action apply --env-file .env --maintenance-confirmed
powercontext server atomic-memory-migrate --action verify --env-file .env
`
Resolve all reported blockers and restart only when verification returns ready: true`. Rollback requires the complete pre-migration backup. The limited unified migration bundle cannot replace this Atomic Memory migration. Reopen Agent sessions after upgrading.
Compatibility changes: retained legacy Memory routes return Atomic records and references, rather than collection revisions or MemoryCitation. Collection Create/Replace, entry revise/retire, capacity, and compaction are no longer supported. Update Memory evidence-gate configuration to declare the policy mode and privacy boundary; do not assume that enabling the gate permits external model calls. Old collection capacity/compaction settings do not constrain Atomic Memory. Custom extraction prompts and unresolved legacy candidates may block migration.
See migration, Atomic Memory compatibility, Artifact search, and Dify.
[Ful
PowerMem v1.1.72026-06-30
What's Changed
* fix(ci): dispatch publish workflow without git checkout by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1116
* feat(plugin): enable transcript inference by default; remove recall/remember skills by @lightzt99 in https://github.com/oceanbase/powermem/pull/1126
* Fix powermem-mcp help to exit without starting the server by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1124
* fix(ci): avoid duplicate generated release notes by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1123
* chore: bump version to 1.1.7 by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1127
Full Changelog: https://github.com/oceanbase/powermem/compare/v1.1.6...v1.1.7
PowerMem v1.1.62026-06-29
What's Changed
* feat: sqlite default to coding agent by @lightzt99 in https://github.com/oceanbase/powermem/pull/1031
* docs: update LOCOMO benchmark results by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1089
* Add Chinese docs and website localization by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1090
* fix(mcp): preserve Snowflake ID precision across MCP tool boundary by @lightzt99 in https://github.com/oceanbase/powermem/pull/1100
* feat: use uv instead of bare python command by @yihong0618 in https://github.com/oceanbase/powermem/pull/1106
* feat(init): AskUserQuestion interactive flow + remote-init mode + dashboard 0.0.0.0 by @lightzt99 in https://github.com/oceanbase/powermem/pull/1101
* fix(scripts): preserve doc dirs in sync-website-docs (Fixes #1096) by @kic635 in https://github.com/oceanbase/powermem/pull/1097
fix(server): push runid filter to storage for session timeline by @kic635 in https://github.com/oceanbase/powermem/pull/1099
* Publish individual release binary assets by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1108
* fix(intelligence): integrate retention fields into runtime paths by @knqiufan in https://github.com/oceanbase/powermem/pull/1084
* fix(intelligence): keep fresh zero-importance memories searchable by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1113
* chore: bump version to 1.1.6 by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1114
New Contributors
* @yihong0618 made their first contribution in https://github.com/oceanbase/powermem/pull/1106
* @kic635 made their first contribution in https://github.com/oceanbase/powermem/pull/1097
Full Changelog: https://github.com/oceanbase/powermem/compare/v1.1.5...v1.1.6
PowerMem v1.1.52026-06-23
What's Changed
* Improve Claude Code plugin init flow by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1024
* fix: clean up graph search tools and results by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1012
* [uv] Use uvx for Claude plugin backend init by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1030
* fix: make graph add relationship writes transactional by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1011
* Fix unit test regressions by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1033
* fix: stabilize regression server restarts by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1034
* build(deps-dev): bump vite from 7.3.2 to 7.3.5 in /dashboard by @dependabot[bot] in https://github.com/oceanbase/powermem/pull/1035
* build(deps): bump launch-editor from 2.12.0 to 2.14.1 in /docs/website by @dependabot[bot] in https://github.com/oceanbase/powermem/pull/1036
* build(deps): bump ws in /docs/website by @dependabot[bot] in https://github.com/oceanbase/powermem/pull/1037
* fix: harden regression server restart handling by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1038
* Support no-LLM mode by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1026
feat: add NoopEmbedding provider for EMBEDDINGPROVIDER=none by @lightzt99 in https://github.com/oceanbase/powermem/pull/1058
* [codex] Add isolated no-LLM Claude hook regression tests by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1050
* feat: move embedding model download into PowerMem with CN-aware routing by @lightzt99 in https://github.com/oceanbase/powermem/pull/1057
* build(deps): bump webpack-dev-server from 5.2.4 to 5.2.5 in /docs/website by @dependabot[bot] in https://github.com/oceanbase/powermem/pull/1064
* Add Claude Code hook privacy controls by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1054
* fix(graph): add UNIQUE index and upsert for entity dedup (#1003) by @jfeng18 in https://github.com/oceanbase/powermem/pull/1032
* Add coding-agent observation and session timeline APIs by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1051
* Merge MCP dependencies into server extra by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1071
* docs: improve .env.example.full reranker and embedding provider documentation by @knqiufan in https://github.com/oceanbase/powermem/pull/1073
fix(intelligence): replace unsafe number-grabbing fallback in parseimportanceresponse by @knqiufan in https://github.com/oceanbase/powermem/pull/1075
* Publish release plugin and platform binaries by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1060
* fix: resolve docs dependency and checkout deprecation by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1079
* chore: bump version to 1.1.5 by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1083
Full Changelog: https://github.com/oceanbase/powermem/compare/v1.1.4...v1.1.5
PowerMem v1.1.42026-06-13
What's Changed
* fix: handle multimodal image failures and native filter operators by @wayyoungboy in https://github.com/oceanbase/powermem/pull/982
* Fix SDK add when embedded seekdb is owned by server by @wayyoungboy in https://github.com/oceanbase/powermem/pull/973
* Expose filters on AgentMemory retrieval APIs by @wayyoungboy in https://github.com/oceanbase/powermem/pull/976
* ci: tolerate missing regression junit artifacts by @wayyoungboy in https://github.com/oceanbase/powermem/pull/999
* build(deps): bump shell-quote from 1.8.3 to 1.8.4 in /docs/website by @dependabot[bot] in https://github.com/oceanbase/powermem/pull/1000
* doc(claude-code-plugin): improve setup auto-detect and dashboard by @lightzt99 in https://github.com/oceanbase/powermem/pull/994
* Fix Node 20 CI deprecation and VS Code extension cache path by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1005
* fix(intelligence): correct Ebbinghaus decay parameters and lifecycle logic by @jfeng18 in https://github.com/oceanbase/powermem/pull/995
* feat(server): auto-open Dashboard in browser on powermem-server start by @knqiufan in https://github.com/oceanbase/powermem/pull/963
* fix: apply access reinforcement to memory decay by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1006
* fix(skill-store): fix broken embed calls and dedup threshold by @jfeng18 in https://github.com/oceanbase/powermem/pull/1001
* fix(dashboard): show empty states for overview charts by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1007
* docs: add Why PowerMem section by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1009
* fix: make lint usable for critical checks by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1008
* docs: add AgentScope MCP integration guide by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1013
* feat(api): filter listed memories by scope by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1015
* fix(init): include server,seekdb extras and improve LLM auto-detection by @lightzt99 in https://github.com/oceanbase/powermem/pull/1016
* fix: scope graph entities by user by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1010
* Fix search score and threshold handling by @wayyoungboy in https://github.com/oceanbase/powermem/pull/1020
* feat(init): change default data dir to ~/.powermem by @lightzt99 in https://github.com/oceanbase/powermem/pull/1019
New Contributors
* @jfeng18 made their first contribution in https://github.com/oceanbase/powermem/pull/995
Full Changelog: https://github.com/oceanbase/powermem/compare/v1.1.3...v1.1.4